EN

CIVITAI / PUBLIC LIBRARY

AI-KSK Civitai 模型与工作流

集中检索 AI-KSK 在 Civitai 公开发布的模型、LoRA 与工作流，查看版本、基础模型和官方来源。

385

公开条目

23

基础模型方向

60732

累计下载

LTX 2.3 Video Control & HD Enhancement Workflow
Workflows 2026-05-13

LTX 2.3 Video Control & HD Enhancement Workflow

This workflow is designed for LTX 2.3 video control and high-definition enhancement. Its main purpose is to take a video or image-guided video input, preserve the original motion structure, and enhance the final result through LTX 2.3 generation, control preprocessing, latent upscaling, audio-video latent routing, and tiled decoding. It is built for creators who want a cleaner, sharper, more stable LTX 2.3 video output instead of a rough low-resolution generation pass. The workflow uses LTX 2.3 as the main video generation backbone, with ltx-2.3-22b-dev_transformer_only_fp8_scaled as the core model route. It also includes Gemma-style LTX text encoding, LTX23 video VAE, LTX23 audio VAE, LTXVPreprocess, EmptyLTXVLatentVideo, LTXVEmptyLatentAudio, LTXVConcatAVLatent, LTXVSeparateAVLatent, SamplerCustomAdvanced, LTXVLatentUpsampler, VAEDecodeTiled, LTXVAudioVAEDecode, and final video output logic. This makes the graph more advanced than a simple image-to-video or video-to-video workflow because it is structured around both control and enhancement. A major strength of this workflow is its video-control preparation section. The graph includes VideoHelperSuite video information reading, frame rate extraction, frame count handling, image resizing, and LTX preprocessing. It also includes optional control preprocessing routes such as DepthCrafter, Canny edge extraction, and DW pose preprocessing. These control modules are useful when the creator wants the generated result to follow the source video’s structure, depth, body movement, edge layout, or camera rhythm more closely. The workflow is also designed for HD improvement. Instead of decoding only the first latent result, it uses LTXVLatentUpsampler and additional refinement sampling stages to push the video toward a higher-quality output. This helps improve detail density, texture clarity, subject sharpness, and final frame polish. The tiled VAE decoding route is important here because high-resolution video decoding can easily become memory-heavy or unstable. Tiled decoding allows the workflow to decode large frames more safely and cleanly. The audio-video route is another practical part of the graph. Audio latent and video latent are connected, separated, decoded, and preserved through the generation process. This makes the workflow suitable for real video production rather than silent visual testing. For AI s

LTXV 2.3 253 downloads
查看公开资料
Anima Preview3 | Tiled Upscale & High-Resolution Detail Refinement Workflow
Workflows 2026-05-13

Anima Preview3 | Tiled Upscale & High-Resolution Detail Refinement Workflow

This workflow is designed for Anima Preview3 tiled upscaling and high-resolution anime detail refinement. Its main purpose is to take an existing anime-style image, enlarge it, split it into manageable tiles, refine each local region with Anima Preview3, and then reconstruct the image into a cleaner high-resolution result. Instead of relying only on a basic pixel upscaler, this workflow combines traditional upscale, tile processing, prompt-assisted regeneration, VAE encoding / decoding, and final image assembly to produce a more polished output. The workflow uses anima-preview3-base.safetensors as the main generation model, qwen_3_06b_base.safetensors as the text encoder, and qwen_image_vae.safetensors as the VAE. It also includes an optional LoRA route, with Gundam_RX78_Flux used in the example setup, showing that the workflow can be adapted for mecha, armor, character, or specific style enhancement tasks. The initial enlargement stage uses 4x_NMKD-Siax_200k.pth, which helps expand the image before the Anima refinement stage begins. The core structure is a tiled enhancement pipeline. The source image is loaded, upscaled, resized to a larger total pixel target, and then divided into tiles through TTP_Tile_image_size and TTP_Image_Tile_Batch. Each tile becomes an independent local region that can be encoded, refined, decoded, and later reassembled. This is important because large anime images often lose detail or become unstable when processed in one full-frame pass. Tiled processing lets the workflow focus more generation attention on each region, improving local texture, line clarity, small mechanical details, clothing folds, hair strands, and background elements. The workflow also uses WD14 tagging. The tile or image content can be analyzed automatically, and the extracted tags can be used to help guide the refinement prompt. This makes the workflow useful when the creator wants the upscaler to understand what is in the image, not just enlarge it mechanically. For anime and illustration work, this can help keep character features, objects, and style direction more coherent during the enhancement stage. The sampling route uses ClownsharKSampler_Beta with a controlled denoise value. This is critical for tiled upscale work. If denoise is too high, each tile may redraw too aggressively and break consistency across the image. If denoise is too low, the imag

Anima 171 downloads
查看公开资料
Anima Preview3 | Local Redraw & Inpainting Workflow
Workflows 2026-05-13

Anima Preview3 | Local Redraw & Inpainting Workflow

This workflow is designed for Anima Preview3 inpainting and local redraw, focusing on controlled partial editing inside an existing anime image. Its main purpose is to let creators mask a specific region, describe the desired change, and regenerate only that local area while keeping the rest of the image stable, coherent, and visually consistent. The workflow uses anima-preview3-base.safetensors as the main generation model, qwen_3_06b_base.safetensors as the text encoder, and qwen_image_vae.safetensors as the VAE. It combines image loading, mask input, mask expansion, blurred mask blending, latent encoding, Differential Diffusion, prompt generation, sampling, VAE decoding, preview, and final saving into one practical local-editing pipeline. Compared with a normal image-to-image workflow, this setup is more suitable when you do not want to redraw the whole image. You only want to modify one selected area. The key part of this workflow is the inpainting mask route. The source image is loaded with its mask, then passed through image_scale_pixel_v2 to normalize the image and mask into a model-friendly size. The mask is then processed by GrowMaskWithBlur, which expands and softens the selected area. This is important because hard mask edges often create visible seams, while a slightly expanded and blurred mask gives the model enough transition space to blend the new content naturally into the original image. The workflow then uses VAEEncode to convert the source image into latent space and SetLatentNoiseMask to restrict the repainting area. This means the sampler is guided to focus on the masked region instead of freely changing the entire frame. The DifferentialDiffusion node further strengthens this local redraw behavior, making the workflow more suitable for precise anime inpainting, object replacement, facial adjustment, hairstyle changes, clothing fixes, prop replacement, and small composition corrections. Another important feature is the built-in prompt-generation logic. The workflow includes an LLM API prompt engineer node, using a role instruction specialized for anime inpainting and partial redraw. It analyzes the reference image, understands the overall composition, art style, lighting, pose, environment, and visual tone, then converts the user’s short edit instruction into a single English prompt line. This is useful because local redraw prompts n

Anima 114 downloads
查看公开资料
Anima Preview3 | Image-to-Image Anime Refinement Workflow
Workflows 2026-05-13

Anima Preview3 | Image-to-Image Anime Refinement Workflow

This workflow is designed for Anima Preview3 image-to-image generation, focusing on controlled anime-style transformation from an existing reference image. Its main purpose is to let creators upload a source image, guide it with a text prompt, and generate a cleaner Anima Preview3 result while still preserving the basic structure, pose, composition, and subject direction of the original picture. The workflow uses anima-preview3-base.safetensors as the main generation model, qwen_3_06b_base.safetensors as the text encoder, and qwen_image_vae.safetensors as the VAE. This creates a compact Anima Preview3 img2img pipeline where the input image is first resized, encoded into latent space, edited through the sampler, decoded back into an image, and then previewed or saved. Compared with a pure text-to-image workflow, this setup gives creators a stronger visual anchor because the model is not starting from an empty latent. It starts from the uploaded image and modifies it according to the prompt. The image preparation section is simple and practical. The source image is loaded through LoadImage, then passed into image_scale_pixel_v2, with the total pixel target set around 1 megapixel and alignment set to 64. This helps normalize the image into a model-friendly size before VAE encoding. The workflow is therefore useful for taking an existing AI draft, sketch, screenshot, character image, animal image, or rough composition and pushing it into a more polished Anima Preview3 style. The workflow then uses VAEEncode to convert the scaled image into latent space. This is the key difference from text-to-image. In text-to-image, the model creates everything from noise. In image-to-image, the original image becomes the base latent, so the final result can keep more of the original layout. This makes it especially useful for redraws, style refinement, concept cleanup, character reinterpretation, and controlled anime transformation. The sampling stage uses ClownsharKSampler_Beta with a 30-step setup, beta57 scheduler, linear/euler sampler route, CFG around 3, and denoise around 0.73. That denoise value is important: it is strong enough to visibly transform the image, but still keeps the source image as a meaningful reference. Lower denoise would preserve the original more strictly; higher denoise would push the result closer to a full redraw. The example prompt is simple:

LTXV 2.3 74 downloads
查看公开资料
LTX 2.3 Distill 1.1 + VBVR 240K | High-Probability Digital Human Workflow
Workflows 2026-05-12

LTX 2.3 Distill 1.1 + VBVR 240K | High-Probability Digital Human Workflow

V This workflow is designed for high-probability LTX 2.3 digital human generation, built around LTX 2.3 Distill 1.1 and VBVR 240K enhancement. Its main purpose is to create a more reliable image-to-video talking-person or digital-avatar result, where the character keeps a stable identity, controlled facial motion, cleaner body movement, and stronger final visual quality. The workflow uses an LTX 2.3 video generation structure with ltx-2.3-22b-distilled-1.1, distilled LoRA support, VBVR-style image-to-video enhancement, Gemma-based LTX text encoding, LTX video VAE, LTX audio VAE routing, NAG enhancement, IC LoRA motion-track control, spatial latent upscaling, custom sampling, tiled decoding, and final video export. This makes it more production-oriented than a simple first-frame animation workflow. The core advantage of this setup is probability and stability. Digital human generation is not only about making a still image move. The model must preserve the face, avoid identity drift, maintain the original clothing and composition, keep the speaking performance believable, and avoid random body motion. This workflow is designed to improve those weak points by using the input image as a strong visual anchor, then reinforcing the generation through LTXVImgToVideoConditionOnly, LTXVPreprocess, audio/video latent routing, and multiple refinement stages. The workflow includes a dedicated audio path. Audio can be encoded through LTXVAudioVAEEncode and connected into the video latent process, allowing the output to behave more like a digital human video rather than a silent image animation. This is useful for AI presenters, talking avatars, product explanation videos, character narration, short drama dialogue, virtual influencer clips, and commercial-style social media content. NAG enhancement is another important part of the workflow. It helps strengthen generation control and reduce unwanted drift during sampling. For digital human videos, this is especially useful because even small errors in the face, mouth, hands, or camera motion can make the result feel unstable. The workflow also uses a motion-track control LoRA to guide the movement more deliberately, helping the character perform with more controlled motion instead of random animation. The pipeline is also staged for better final quality. It first builds the base video from the input image and conditi

LTXV 2.3 187 downloads
查看公开资料
WAI-ANIMA Image-to-Image | Controlled Anime Style Refinement Workflow
Workflows 2026-05-12

WAI-ANIMA Image-to-Image | Controlled Anime Style Refinement Workflow

This workflow is designed for WAI-ANIMA image-to-image generation and controlled anime-style refinement. Its main purpose is to take an existing image as the visual foundation, then use the WAI-ANIMA model pipeline to redraw, enhance, stylize, or push the image toward a cleaner anime key visual while still keeping the original composition as an anchor. Compared with a pure text-to-image workflow, this setup gives creators more control because the input image is not discarded. The uploaded image is loaded into the workflow, resized through a pixel-scaling node, encoded into latent space through the Qwen image VAE, and then processed through the WAI-ANIMA generation model. This means the final result is guided by both the source image and the written prompt, making it useful for style conversion, anime redraws, character enhancement, composition preservation, and prompt-based visual refinement. The workflow uses waiANIMA_v10 as the main UNet model, qwen_3_06b_base as the text encoder, and qwen_image_vae as the VAE. This combination is aimed at anime-style image generation with stronger prompt understanding and cleaner visual rendering. The workflow is compact, direct, and easy to modify, making it suitable for creators who want a simple but practical Anima image-to-image setup instead of a large multi-stage graph. The input image is scaled to around a 1-megapixel working size through image_scale_pixel_v2. This helps normalize the image before it enters the latent process, avoiding extremely small or oversized inputs that may reduce stability. After that, VAEEncode converts the image into latent space, allowing the sampler to modify the image with controlled denoise rather than generating from a blank latent. The sampler section uses ClownsharKSampler_Beta with a 30-step beta-style sampling route, CFG control, and a moderate denoise setting. This is important for image-to-image work. If denoise is too low, the result may barely change. If denoise is too high, the original structure can be lost. The setup here is meant to balance source-image preservation with visible anime-style transformation. The positive prompt in the workflow is built around a high-quality anime key visual: adult anime beauty, Anima style, fantasy atmosphere, cinematic composition, detailed clothing, dramatic scale, and polished illustration finish. The negative prompt suppresses commo

LTXV 2.3 169 downloads
查看公开资料
Prompt Relay Director | LTX 2.3 Dual-Character LoRA Dialogue Workflow
Workflows 2026-05-12

Prompt Relay Director | LTX 2.3 Dual-Character LoRA Dialogue Workflow

This workflow is designed as a Prompt Relay director for LTX 2.3 storytelling, dual-character dialogue, and dual-character LoRA prompt formatting. Its main purpose is not to generate the video directly, but to help creators turn messy ideas, rough Chinese notes, images, storyboard grids, short dialogue drafts, or incomplete scene concepts into a clean LTX 2.3 Prompt Relay format using only two fields: global_prompt and local_prompts. The workflow uses an LLM API node as the prompt director. It can analyze uploaded images, multiple images, rough prompts, short Chinese phrases, dialogue ideas, style requests, duration requirements, no-dialogue instructions, and storyboard-style references. After analyzing the input, it restructures everything into a usable Prompt Relay prompt that is much easier to paste into an LTX 2.3 video workflow. The core value of this workflow is standardization. Many LTX 2.3 video failures are not caused by the model alone, but by prompts that mix global rules, timeline events, dialogue, camera instructions, sound design, and negative restrictions in the wrong place. This workflow separates those elements clearly. The global_prompt locks the full-video anchors: scene, character identities, left-right placement, costumes, props, lighting, camera baseline, sound atmosphere, dialogue status, and stability restrictions. The local_prompts only describe time-specific events: actions, gestures, reactions, camera movement, object changes, dialogue lines, and timed sound effects. This is especially useful for dual-character LoRA scenes. Two-person video generation is difficult because characters may swap sides, merge, change style, lose identity, or produce chaotic body movement. This workflow forces the prompt to lock left-right positions, stable identities, stable character styles, natural eye contact, controlled gestures, and medium two-shot camera logic. It also adds restrictions such as no face swapping, no duplicated characters, no extra people, no wrong mouth movement, no mismatched lip sync, no subtitles, no screen text, no watermark, no jump cuts, and no random costume or style changes. The workflow also handles dialogue intelligently. If the user provides dialogue, the prompt director shortens and rewrites it into natural Mandarin lines suitable for lip sync. If the user does not ask for dialogue, it does not force speaking. If th

LTXV 2.3 151 downloads
查看公开资料
JoyAI Image vs Qwen 2511 | Single-Image Editing Comparison Workflow
Workflows 2026-05-12

JoyAI Image vs Qwen 2511 | Single-Image Editing Comparison Workflow

This workflow is designed for comparing JoyAI-Image single-image editing with Qwen Image Edit 2511. Its main purpose is to help creators test how different image-editing routes handle the same or similar visual editing task, especially when the goal is not only to change an image, but also to preserve identity, structure, lighting, texture, and final visual quality. The workflow contains a JoyAI-Image editing branch and a Qwen Image Edit 2511 comparison branch. The JoyAI branch uses JoyAI-Image-Und-merger_bf16 as the visual understanding / CLIP route, joy_image_transformer as the main image transformer, Wan2.1_VAE.pth as the VAE, JoyAI_Image_ENCODER for image-and-prompt conditioning, JoyAI_Image_LATENTS for latent preparation, JoyAI_Image_SM_KSampler for sampling, and JoyAI_Vae_Decoder for final decoding. This route is focused on direct single-image instruction editing: upload one image, write the edit command, and let the model transform the image while keeping the important visual identity. The Qwen 2511 branch uses qwen_image_edit_2511_bf16 with Qwen Image Edit 2511 Lightning 4-step LoRA, qwen_2.5_vl_7b_fp8_scaled as the visual-language encoder, qwen_image_vae, TextEncodeQwenImageEditPlus, reference-latent method nodes, KSampler, and tiled VAE decoding. This branch is useful for comparing how Qwen 2511 handles detail, lighting, identity retention, reference-image understanding, and speed under a lightweight accelerated setup. The workflow is especially useful for creators who want to answer practical editing questions: Which route keeps the original character identity better? Which one follows text instructions more directly? Which one changes the background more naturally? Which one produces better lighting and shadow integration? Which one is faster for batch testing? Which one is better for Civitai preview images, RunningHub demos, social media covers, or local production? The JoyAI example focuses on a blue-haired cyber mechanical girl being transferred into a bright grassland scene while preserving her facial features, twin-tail hairstyle, mechanical ear devices, neck structure, and metallic body. The prompt also asks for realistic outdoor lighting, grassland atmosphere, natural shadows, metal highlights, and reduced cutout feeling. This makes it a strong test for identity preservation and scene replacement. The Qwen 2511 side includes a portrai

LTXV 2.3 87 downloads
查看公开资料
LTX 2.3 Subtitle Remover | AI Video Cleanup 0.5 Workflow
Workflows 2026-05-12

LTX 2.3 Subtitle Remover | AI Video Cleanup 0.5 Workflow

This workflow is designed for LTX 2.3 video subtitle removal and visual text cleanup, built as a reference 0.5 version for repairing unwanted text pollution inside AI-generated or authorized video materials. Its main purpose is to help creators remove subtitles, captions, random text artifacts, watermark-like overlays, and other unwanted visual marks from a video while keeping the original scene, motion, lighting, and unmasked areas as stable as possible. Unlike a simple crop, blur, or cover-up method, this workflow is based on mask-guided video repair. The unwanted subtitle or text region is isolated through a mask pipeline, then the model reconstructs the damaged area using the surrounding visual context. This makes the result more natural because the repaired area is regenerated to match the original background, rather than being hidden by a flat patch or blurred block. The workflow uses an LTX 2.3 repair route with video VAE, audio VAE, LTX conditioning, custom sampling, mask processing, and audio-video export logic. It includes structured SetNode / GetNode routing for base model, video VAE, audio VAE, CLIP, FPS, final masks, audio, and resolution management. This makes the graph more modular and easier to reuse in a production environment, especially when the user needs to repeatedly process different videos with similar subtitle or text-contamination problems. A key part of this workflow is the mask preparation section. The workflow includes mask handling tools such as BlockifyMask, final mask routing, and latent noise mask logic. This matters because video subtitle repair depends heavily on the mask quality. If the mask is too small, the text may remain. If the mask is too large, the model may unnecessarily change clean background areas. A good mask should cover the unwanted text fully while preserving enough surrounding context for the model to rebuild the background naturally. The workflow also keeps the audio route in the graph. Audio can be carried through the pipeline and reattached to the final output, which makes the workflow more practical for actual video repair instead of isolated frame testing. The graph includes audio retrieval, audio trimming / duration logic, LTXVAudioVAEEncode, LTXVConcatAVLatent, and final video creation. This allows the repaired result to remain usable as a complete video output. The sampling route uses LTX 2.3 v

LTXV 2.3 80 downloads
查看公开资料
LTX 2.3 Dual-Character Dialogue | Enhanced Cinematic I2V Workflow
Workflows 2026-05-12

LTX 2.3 Dual-Character Dialogue | Enhanced Cinematic I2V Workflow

This workflow is designed for LTX 2.3 dual-character dialogue video generation, focusing on stable two-person interaction, controlled left-right positioning, stronger motion continuity, and more polished final video quality. Its main purpose is to help creators turn a single reference image into a cinematic two-character conversation shot, where both characters remain visually consistent and the scene keeps a coherent dialogue rhythm instead of drifting into random motion. The workflow uses LTX 2.3 as the main video generation backbone, with ltx-2.3-22b-distilled-1.1 as the core checkpoint route. It also uses Gemma-style LTX text encoding, LTX audio VAE routing, LTX video VAE decoding, image-to-video conditioning, NAG enhancement, VBVR I2V LoRA support, latent upscaling, multi-stage sampling, and final video export. This makes it more advanced than a basic I2V workflow because it is built specifically for controlled character interaction rather than simple first-frame animation. The core strength of this workflow is two-character stability. In many AI video workflows, two-person scenes are difficult because characters may swap positions, merge into each other, lose identity, change clothing, or break the left-right relationship. This workflow is designed to reduce those problems by using the input image as a strong visual anchor, then reinforcing the generation with LTXVImgToVideoConditionOnly, LTXVConditioning, and a structured prompt. The negative prompt also directly suppresses subtitles, scene cuts, glitches, warping, extra limbs, extra hands, static frames, low-quality motion, and unwanted transitions. Another important feature is the VBVR I2V LoRA route. The workflow loads an LTX 2.3 VBVR I2V LoRA, which helps strengthen image-to-video behavior, motion consistency, and prompt adherence. This is especially useful for dialogue-style videos, where small gestures, facial direction, eye contact, and body positioning matter more than large chaotic movement. The workflow also includes NAG enhancement. NAG is used to improve guidance stability and reduce generation drift during sampling. For dual-character dialogue scenes, this matters because the video must preserve not only the scene style, but also the relationship between the two characters. The left character should remain on the left, the right character should remain on the right, and both should co

LTXV 2.3 91 downloads
查看公开资料
VBVR Digital Human | Stable Talking Avatar Workflow
Workflows 2026-05-12

VBVR Digital Human | Stable Talking Avatar Workflow

This workflow is designed for VBVR digital human video generation, focusing on stable talking-avatar animation from a single reference image and an audio input. Its main purpose is to help creators generate a more controlled digital human result where the character keeps the same face, framing, camera angle, clothing, and background while performing natural speaking motion. The workflow is built around an LTX 2.3 video generation pipeline with VBVR I2V LoRA enhancement, LTX audio / video latent routing, Gemma-style text encoding, LTX video VAE, LTX audio VAE, NAG enhancement, IC LoRA motion-track control, spatial latent upscaling, multi-stage sampling, tiled decoding, and final video export. Compared with a basic image-to-video workflow, this setup is more suitable for talking-avatar production because it combines visual anchoring, audio routing, and controlled motion guidance in one graph. The core idea is simple but important: keep the shot stable and make the woman speak. In digital human generation, the biggest problem is often not whether the image can move, but whether it moves too much. A weak workflow may change the face, zoom the camera out, alter the clothing, deform the mouth, create extra hands, or shift the background. This workflow is designed to reduce those problems by keeping the camera and scene steady while concentrating motion on the face, mouth, head, and subtle body performance. VBVR is used here as an image-to-video consistency and motion-control booster. It helps the model follow the source image more closely and reduces random drift during generation. This is especially important for digital human videos because the first frame usually defines the person’s identity. If the generated video loses that identity after a few seconds, the result becomes unusable for avatar content, product narration, AI presenters, or character dialogue. The workflow also includes an audio latent route. The audio is encoded through LTXVAudioVAEEncode, connected into the audio/video latent structure, and later separated and decoded for final output. This makes the workflow more than a silent animation setup. It is designed for speaking-person videos where the final result needs both visual motion and usable audio-video export. Another important part is the use of NAG and IC LoRA motion-track control. NAG helps stabilize generation guidance, while the m

LTXV 2.3 70 downloads
查看公开资料
LTX 2.3 Four-Image Reference Video Workflow
Workflows 2026-05-11

LTX 2.3 Four-Image Reference Video Workflow

This workflow is designed for LTX 2.3 four-image reference video generation, allowing creators to build a continuous cinematic video from multiple visual references instead of relying on a single image or a text prompt alone. Its main purpose is to use four reference images as visual anchors for scene design, character appearance, product details, atmosphere, and motion continuity, then generate a coherent LTX 2.3 video with stronger control over identity, environment, and visual logic. The workflow uses LTX 2.3 Dev-Dare merged distilled components as the main video generation backbone, together with Gemma 3 text encoding, LTX audio / video latent routing, LTX video VAE decoding, and a custom sampler structure. It also includes enhancement routes such as LTX2.3 Crisp Enhance and VBVR I2V-style LoRA support, making it more suitable for controlled image-to-video generation where visual consistency matters. The key feature is the multi-reference guide structure. Four images can be loaded and processed separately, then resized and prepared through ImageResizeKJv2 and LTXVPreprocess. These references are injected into the generation process through LTXVAddGuideMulti. This gives the model more than one visual source to follow. One image can define the main subject, another can define the product, another can provide the scene or lighting mood, and another can guide the final visual style or atmosphere. The workflow also uses PromptRelayEncodeTimeline, which is important for video control. Instead of writing only one static prompt, the workflow can describe the video as a time-based sequence. The global prompt defines the overall scene, subject, atmosphere, and visual rules, while local prompts describe what happens in different time segments. This makes the final video more cinematic and less random, because the model receives both a stable visual concept and a clear temporal direction. In the uploaded setup, the workflow is structured around a polished beauty product showcase. The prompt describes a young adult East Asian woman presenting a blue skincare serum bottle in a clean studio or cinematic environment, with controlled camera movement, soft light, product visibility, and stable character performance. This shows the intended strength of the workflow: it is not only for abstract animation, but also for commercial-style video generation, product demonstra

LTXV 2.3 404 downloads
查看公开资料
Anima Preview2 Tiled Upscale Workflow
Workflows 2026-05-11

Anima Preview2 Tiled Upscale Workflow

This workflow is designed for Anima Preview2 tiled upscaling and detail enhancement. Its main purpose is to take an existing image, enlarge it to a higher usable resolution, then refine the enlarged result tile by tile with Anima Preview2 so the final image looks cleaner, sharper, and more detailed without relying on simple interpolation alone. It is especially useful for creators who already have a good base image but want a higher-quality final version for publishing, preview display, cover images, or showcase output. The workflow uses anima-preview2.safetensors as the main refinement model, qwen_3_06b_base.safetensors as the text encoder, and qwen_image_vae.safetensors as the VAE. It also includes a LoRA loader in the graph, showing that the upscaling route can be combined with an additional style or detail bias when needed. The overall design is not just “make the image bigger.” Instead, it builds a multi-stage pipeline: upscale first, split into tiles, describe the tiles, refine them in latent space, decode them, and then reassemble the final image. The process starts with a source image loaded into the workflow. That image is first enlarged through a traditional upscale model using 4x_NMKD-Siax_200k. After that, the result is further normalized with ImageScaleToTotalPixels, targeting a larger working size while keeping the image manageable. This gives the workflow a stronger high-resolution base before diffusion refinement begins. A major feature of this workflow is tiled processing. The enlarged image is divided into tiles with TTP_Image_Tile_Batch, and the tile layout is controlled through TTP_Tile_image_size. This is important because very large images can be difficult to refine in a single pass, especially when you want detail recovery without destroying the whole composition. By splitting the image into tiles, the workflow can enhance local detail more effectively while still reconstructing the whole image afterward through TTP_Image_Assy. Another useful feature is automatic tile captioning. The workflow uses WD14Tagger to analyze the image tiles and generate prompt-like tag information. Those generated tags are passed through ShowText and then into the positive CLIPTextEncode route. This means the workflow does not depend entirely on a manually written prompt. Instead, it can derive a descriptive prompt from the source content itself, which h

Anima 218 downloads
查看公开资料
Anima Preview2 Image-to-Image Workflow
Workflows 2026-05-11

Anima Preview2 Image-to-Image Workflow

This workflow is designed for Anima Preview2 image-to-image generation, giving creators a clean and efficient way to transform an input image into a more polished anime-style result while still preserving the original composition and core visual structure. Unlike a pure text-to-image workflow, this setup begins with a source image, analyzes it, converts it into latent space, and then regenerates it through Anima Preview2 with prompt guidance. This makes it especially useful for style transfer, anime enhancement, visual cleanup, character refinement, and turning an existing image into a more cinematic anime illustration. The workflow uses anima-preview2.safetensors as the main generation model, qwen_3_06b_base.safetensors as the text encoder, and qwen_image_vae.safetensors as the VAE. It starts by loading a source image, then rescales it with image_scale_pixel_v2 so the input is normalized to a model-friendly pixel target while keeping alignment stable. After that, the image is encoded into latent space with VAEEncode, which becomes the structural starting point for the generation stage. This design makes the workflow ideal for users who want to preserve the general framing, subject position, and image layout rather than generating from scratch. One of the useful features in this workflow is the WD14 tagger support. The loaded image is automatically analyzed to produce tag-style prompt information, which can then be combined with the manual positive prompt. In the provided setup, the base prompt includes a beach-sunset singing-girl concept, showing how this workflow can mix image-derived tags with user-written prompt direction. This is practical for image-to-image generation because it reduces prompt-writing difficulty and helps the model better understand the existing content of the input image. The negative prompt is focused on common quality issues, including low quality, blur, bad anatomy, bad hands, extra fingers, fused fingers, deformed faces, text, watermark, logo, and JPEG artifacts. This helps keep the result clean during regeneration, especially when the source image already contains imperfect details or when the user wants a sharper anime-style finish. The main generation stage uses a KSampler with 40 steps, CFG around 3, DPM++ 2M SDE GPU sampling, SGM Uniform scheduling, and denoise around 0.75. This is important because the denoise value make

Anima 199 downloads
查看公开资料
XiaoYunque Seamless Text Removal | LTX 2.3 Watermark and Subtitle Video Repair Workflow
Workflows 2026-05-11

XiaoYunque Seamless Text Removal | LTX 2.3 Watermark and Subtitle Video Repair Workflow

This workflow is designed for LTX 2.3 video inpainting and authorized video cleanup, focusing on removing unwanted subtitles, watermarks, overlay text, random AI letters, logo marks, and visual text artifacts from a video while keeping the original motion and scene continuity as stable as possible. It is not a simple blur or crop solution. The goal is to reconstruct the damaged area frame by frame so the repaired region blends naturally with the original footage. The workflow uses an LTX 2.3 video repair route with a GGUF-based LTX 2.3 model, LTX23 video VAE, LTX23 audio VAE, LTX conditioning, and a custom sampling structure. It also applies two important LoRA directions: an Edit Anything global LoRA and an inpaint masked R2V LoRA. This combination makes the workflow more suitable for targeted video repair, because the model is guided to understand both the original video context and the masked area that needs to be regenerated. The main logic is mask-based video restoration. The user provides a video and a mask area that covers the unwanted text, subtitle, watermark, or damaged region. The workflow then prepares the video latent, applies the mask as a noise mask, and lets LTX 2.3 regenerate only the target area while preserving the rest of the frame. This is especially important for video repair, because uncontrolled regeneration can easily change the face, background, lighting, camera motion, or scene details outside the repair zone. The prompt and negative prompt are also critical. The negative prompt explicitly suppresses subtitles, captions, Chinese subtitles, watermarks, logos, overlay text, unreadable text, ghost text, flicker, color shift, inconsistent background, blurry patches, and duplicated edges. This helps the workflow understand that the target is clean reconstruction, not adding new text or replacing the scene with unrelated content. The workflow also includes audio-aware routing. Audio can be encoded through the LTX audio VAE and combined with video latent processing, while the final result is exported through VHS_VideoCombine as an MP4 file. This makes the workflow suitable for practical video cleanup instead of isolated frame repair. This setup is useful for repairing your own AI-generated videos, removing accidental prompt text, fixing subtitle contamination, cleaning logo artifacts from authorized material, restoring damaged video a

LTXV 2.3 100 downloads
查看公开资料
LTX 2.3 Three-Image Reference Video Workflow
Workflows 2026-05-11

LTX 2.3 Three-Image Reference Video Workflow

This workflow is designed for LTX 2.3 three-image reference video generation, giving creators a controlled way to turn multiple visual references into a coherent cinematic video. Instead of relying on only one source image, this workflow uses three separate reference images to guide the final result, making it more practical for character consistency, product presentation, scene control, and short-form AI video production. The core idea is multi-reference visual anchoring. A single image often cannot provide enough information for a stable video. It may show the character clearly, but not the product. It may show the lighting, but not the intended camera angle. It may show the scene, but not the subject identity. By using three reference images, this workflow gives the model more visual context. One image can define the main character or subject, the second image can define the product, object, clothing, or key design element, and the third image can provide the background, mood, color palette, or scene atmosphere. This makes the workflow especially useful for commercial-style video generation. For example, creators can use it to build AI influencer clips, beauty product showcases, fashion previews, character-driven advertisements, cinematic product reveals, short social media videos, and Civitai / RunningHub demonstration assets. The prompt can then act as the director, telling the model how the three references should be combined and how the action should develop over time. The workflow is based on an LTX 2.3 video generation route, using image reference guidance, prompt conditioning, video latent creation, sampling, decoding, and final video export. In a typical use case, the reference images are resized and prepared before being passed into the video generation stage. The model then uses these images as guide signals while following the written prompt. This helps the final output stay closer to the intended visual design instead of drifting into a random text-only result. The strength of this workflow is not just reference fusion, but reference fusion for video. In image generation, a reference mismatch may only affect one frame. In video generation, that mismatch can become flicker, identity drift, unstable clothing, object deformation, or inconsistent backgrounds. By giving the workflow three visual anchors, creators can improve the chance that the

LTXV 2.3 224 downloads
查看公开资料
LTX 2.3 Four-Image Reference Audio-Driven Video Workflow
Workflows 2026-05-11

LTX 2.3 Four-Image Reference Audio-Driven Video Workflow

This workflow is designed for LTX 2.3 four-image reference audio-driven video generation. It combines multiple visual references with audio-aware video latent routing, making it suitable for creators who want a more controlled cinematic video instead of a random text-only result. The main purpose is to use four reference images as visual anchors, then guide the video generation with prompt structure, temporal motion planning, and audio-related conditioning so the final output feels more coherent, more rhythmic, and more production-ready. The workflow is built around the LTX 2.3 video generation system, using LTX video and audio latent components, LTX VAE decoding, Gemma-style text conditioning, custom sampler routes, manual sigma control, and final video export. Compared with a simple image-to-video workflow, this setup is more advanced because it does not depend on only one image. It allows the user to provide multiple reference images that can define different parts of the final video: character identity, product appearance, clothing or pose, background atmosphere, color tone, camera style, and visual direction. The four-image reference structure is the most important visual control layer. In practical use, Image 1 can define the main subject, Image 2 can provide the product or object, Image 3 can guide the scene or environment, and Image 4 can provide the final mood, style, or lighting reference. This gives LTX 2.3 more visual information to work with, reducing the chance of identity drift, unstable product appearance, or inconsistent scene design. For product videos, AI influencer clips, fashion showcases, beauty ads, music-video style shots, and short-form commercial content, this kind of multi-reference structure is much more useful than single-image generation. The audio-driven part makes this version different from the normal four-image reference workflow. The graph includes audio VAE routing, audio latent connection, LTXVConcatAVLatent, LTXVSeparateAVLatent, and LTXVAudioVAEDecode-style processing, allowing the video pipeline to carry audio information through the generation and export process. This makes the workflow suitable for videos where rhythm, performance, presentation timing, music atmosphere, or spoken content matters. It is not just a silent image animation pipeline; it is structured for video output with audio-aware handling. The wor

LTXV 2.3 192 downloads
查看公开资料
LTX 2.3 Text Artifact Remover | AI Video Cleanup Workflow
Workflows 2026-05-11

LTX 2.3 Text Artifact Remover | AI Video Cleanup Workflow

This workflow is designed for LTX 2.3 AI video text artifact cleanup, focusing on repairing unwanted subtitles, random AI-generated letters, watermark-like text pollution, overlay captions, ghost text, logo artifacts, and other visual contamination inside video frames. Its main purpose is to help creators clean their own generated videos or authorized materials by reconstructing the damaged area instead of simply blurring, cropping, or covering it. The workflow uses an LTX 2.3 video inpainting route based on a GGUF LTX 2.3 distilled model, LTX23 video VAE, LTX23 audio VAE, Gemma-style text conditioning, custom sampler control, and LoRA-assisted repair. The key repair direction is built around LTX 2.3 Edit Anything and inpaint-style LoRA logic, allowing the model to understand that the masked area should be regenerated while the unmasked region should remain unchanged. The core prompt is very direct: remove the subtitle, watermark, or text inside the masked area, reconstruct the occluded background naturally and seamlessly, and keep the original scene, camera angle, lighting, motion, composition, and all unmasked regions unchanged. This is exactly what makes the workflow useful for AI video cleanup. It is not trying to redesign the whole shot. It is trying to surgically repair the polluted area. The negative prompt is also targeted for this use case. It suppresses subtitles, captions, text, Chinese subtitles, watermarks, logos, overlay text, random letters, unreadable text, ghost text, flicker, color shift, inconsistent background, blurry patches, and duplicated edges. These negative controls are important because text-removal workflows often fail by leaving behind soft stains, repeated edges, or new unreadable letters. This setup tries to reduce those artifacts during regeneration. The workflow also contains an audio latent route, using LTXVAudioVAEEncode and LTXVConcatAVLatent. Even when the repair task is mainly visual, keeping the LTX audio / video latent structure makes the pipeline more suitable for actual video production. The video latent is sampled through SamplerCustomAdvanced, then separated, cropped, decoded, and prepared for output. This makes it closer to a real repair workflow rather than a single-frame test. This setup is useful for fixing AI-generated video mistakes, removing accidental prompt text, cleaning subtitle pollution, repairing

LTXV 2.3 89 downloads
查看公开资料
Qwen Image Edit 2511 Seamless Head Swap Redraw-and-Paste-Back Master Workflow
Workflows 2026-05-10

Qwen Image Edit 2511 Seamless Head Swap Redraw-and-Paste-Back Master Workflow

This workflow is designed for Qwen Image Edit 2511 seamless head replacement with redraw-and-paste-back control. Compared with a simple head-swap workflow, this version adds a more complete local editing structure: the target region is prepared, redrawn with Qwen Image Edit 2511, then pasted back into the original image area with stronger composition preservation. The goal is to replace the head identity naturally while keeping the original body, clothing, background, lighting, camera angle, and image layout stable. The workflow uses Qwen Image Edit 2511 as the main editing route, with qwen_image_vae, qwen_2.5_vl_7b_fp8_scaled, and Qwen 2511-related model / LoRA components. It is structured for controlled image editing rather than full-image regeneration. This is important because head replacement can easily break the original body proportion, neck connection, shoulder alignment, light direction, skin tone, and facial scale. This workflow is designed to reduce those problems through masking, reference preparation, and paste-back reconstruction. The key idea is “redraw first, then restore.” The workflow prepares the main image and reference image, uses background / subject extraction and mask processing to isolate the editable region, then expands and softens the mask with MaskGrow-style controls. After that, Qwen Edit configuration nodes prepare the image references, visual-language inputs, VAE images, masks, and prompt structure. The edited result is decoded, cropped with pad information, scaled back when needed, and restored into the original image through a paste-back / restore-crop-box route. This gives the workflow a more practical production behavior. Instead of allowing the model to reinterpret the entire image, the edit is concentrated on the face / head area. The original image remains the structural anchor, while the reference image provides the replacement identity. This makes the final result more suitable for portrait correction, character testing, fashion model previews, authorized identity replacement, cosplay transformation, AI influencer image editing, and Civitai / RunningHub showcase examples. The workflow also includes preview and comparison tools, such as mask preview, image comparer, image reel display, and before-after outputs. These are useful for checking whether the mask is correct, whether the reference identity is preserved, a

Qwen 300 downloads
查看公开资料
Z-Image-i2L (Image to LoRA) Fast LoRA Training + ControlNet Testing + Upscale Workflow
Workflows 2026-05-10

Z-Image-i2L (Image to LoRA) Fast LoRA Training + ControlNet Testing + Upscale Workflow

This workflow is an expanded Z-Image-i2L production pipeline that combines fast Image-to-LoRA generation, ControlNet structure testing, and high-resolution tiled upscaling into one complete ComfyUI graph. It is designed for creators who do not only want to generate a quick LoRA from reference images, but also want to immediately test that LoRA under real production conditions and then push the result into a more polished final output. The first stage focuses on fast LoRA creation. Multiple reference images are loaded and combined into a training image batch, then passed into the RunningHub Z-Image-i2L system. This allows the workflow to generate a lightweight Z-Image LoRA from a small group of images without requiring a traditional local training setup, dataset folder preparation, caption files, or command-line configuration. It is especially useful for quickly capturing a character identity, fashion style, product look, object concept, creature design, or consistent visual aesthetic. After the LoRA is generated, the workflow immediately saves it and loads it back into Z-Image Base for testing. This is the key advantage of the pipeline: training and validation happen in the same graph. The user can quickly see whether the generated LoRA actually affects the output, whether it preserves the target identity, whether it introduces artifacts, and whether the strength needs to be adjusted. This makes the workflow much more practical than a training-only setup. The second stage adds ControlNet testing. A structure reference image is processed through DepthAnythingV2Preprocessor to create a depth map, then applied through Z-Image Fun ControlNet Union. This lets the newly generated LoRA be tested under controlled composition, depth, layout, and spatial guidance. A LoRA may look fine in a basic text-to-image test, but fail when the camera angle or scene structure becomes more demanding. This workflow helps reveal that immediately. The generation section uses Z-Image Base with qwen_3_4b text encoding, AE VAE, ControlNet guidance, SplitSigmas, DetailDaemonSamplerNode, CFGGuider, and SamplerCustomAdvanced. This gives the workflow a more controlled two-stage sampling structure, where the early phase builds the main layout and the later phase refines the image. It is useful for evaluating prompt compatibility, LoRA strength, structural stability, and final image coher

ZImageTurbo 180 downloads
查看公开资料
Qwen Image Edit 2511 Seamless Head Swap Master Workflow
Workflows 2026-05-10

Qwen Image Edit 2511 Seamless Head Swap Master Workflow

This workflow is designed for Qwen Image Edit 2511 seamless head replacement, focused on controlled, authorized portrait editing with strong identity transfer and natural blending. Its main purpose is to use one image as the body / scene reference and another image as the head / face identity reference, then generate a clean final result where the replacement looks integrated rather than pasted. The workflow is built around Qwen Image Edit 2511, using qwen_image_edit_2511_fp8mixed / qwen_image_edit_2511 model routing, qwen_2.5_vl_7b_fp8_scaled as the vision-language text encoder, and qwen_image_vae as the VAE. It also includes a dedicated head-swap LoRA route, such as bfs_head_v5_2511_merged_version_rank_16_fp16, to strengthen the model’s ability to perform realistic and stable head replacement. This makes the workflow more specialized than a normal image-editing graph. The core idea is not to regenerate the whole image. Instead, the workflow uses inpainting and reference control to focus the edit on the head area while preserving the body, clothing, pose, background, lighting, camera angle, and overall composition from the base image. This is important because many face-swap or head-swap attempts fail by changing the entire person, destroying the original body structure, or creating mismatched lighting and skin tone. This workflow is designed to reduce that drift. The workflow includes InpaintCropImproved and InpaintStitchImproved, which help crop the target region, process the edit, and stitch the result back into the original image. This gives the workflow a more professional local-editing structure. It also uses ReferenceLatent, FluxKontextMultiReferenceLatentMethod, CFGNorm, ModelSamplingAuraFlow, SamplerCustomAdvanced, and VAEDecode to improve reference stability and output quality. A key part of this workflow is mask and foreground preparation. BiRefNetRMBG is included for background removal and subject isolation, helping the system identify the usable human region more clearly. The workflow also uses ImageResizeKJv2, image concatenation, preview nodes, and comparison output, making it easier to check the body image, head reference, generated result, and before-after differences. The prompt logic is written specifically for seamless replacement. It asks the model to preserve the body image environment, framing, lighting, camera perspective, expos

Qwen 299 downloads
查看公开资料
VACE SkyReels V3 R2V Merge Skeleton-Guided Video Workflow
Workflows 2026-05-10

VACE SkyReels V3 R2V Merge Skeleton-Guided Video Workflow

This workflow is designed for VACE + SkyReels V3 R2V Merge skeleton-guided video generation. Its main purpose is to take a reference character or source image, combine it with a motion / pose-guidance video, and generate a new video where the subject follows the skeleton-driven movement while preserving a stronger visual identity and cinematic style. It is especially useful for creators who want more controlled character motion instead of relying only on text prompts or random video generation. The workflow uses a video-loading and motion-reference structure. VHS_LoadVideo imports the guide video, extracts frames, reads video information such as width, height, frame count, and audio, then passes this information into the generation and output stages. This is important because skeleton-guided video workflows need the generated result to follow the timing and motion structure of the input video. The workflow also keeps audio routing available, so the final exported video can preserve the source audio when needed. On the visual side, the workflow uses reference images and image resizing nodes to prepare the character or visual identity before generation. ImageResizeKJv2 and ImageResize+ help align the reference image and guide-video dimensions, making the input more compatible with the video generation route. The workflow also uses image batching and comparison layouts, allowing users to check reference images, generated frames, and side-by-side outputs more clearly. The generation route is built around a VACE / SkyReels-style video pipeline with Wan-family components. It uses UMT5-style text encoding, Wan VAE decoding, positive and negative prompt conditioning, KSampler generation, VAEDecode, and VHS_VideoCombine for final MP4 output. The positive prompt controls the subject, scene, style, and action direction, while the negative prompt helps suppress common video artifacts such as broken limbs, unstable anatomy, flicker, low quality, or inconsistent motion. The key value of this workflow is skeleton-driven motion transfer. Instead of asking the model to invent an action from text alone, the workflow uses the guide video as a motion structure. This makes it more suitable for dance videos, martial arts motion, character performance, action clips, stylized animation, AI influencer videos, cosplay transformation, game-character motion tests, and cinematic sho

Wan Video 14B t2v 87 downloads
查看公开资料
FireRed-Image-Edit-1.0 Dual-Image Editing Workflow
Workflows 2026-05-10

FireRed-Image-Edit-1.0 Dual-Image Editing Workflow

This workflow is designed for FireRed-Image-Edit-1.0 dual-image editing, mainly focused on transferring visual information from one image to another while keeping the target subject stable. A typical use case is: Image 1 provides the main person, pose, face, body structure, and composition, while Image 2 provides clothing, style, texture, or a reference object. The workflow then edits Image 1 according to Image 2 and the text instruction, making it suitable for outfit transfer, style replacement, character restyling, product reference editing, and controlled visual fusion. The workflow is built around FireRed-Image-Edit-1.0_fp8_e4m3fn.safetensors as the main editing model. It uses qwen_2.5_vl_7b_fp8_scaled.safetensors as the Qwen image text encoder and qwen_image_vae.safetensors as the VAE. It also applies Qwen-Image-Lightning-8steps-V2.0 as a model-only LoRA acceleration route, with the LoRA strength set lower than a full-force generation path. This makes the workflow practical for fast dual-image editing while still keeping enough model control for reference-based transformation. The core editing logic uses TextEncodeQwenImageEditPlus, which allows the prompt to be conditioned together with multiple image references. In this workflow, the prompt example is simple and direct: “the woman in Image 1 wears the clothes from Image 2.” This structure is very useful because the user does not need to describe every clothing detail manually. Image 2 provides the visual reference, while the prompt defines the editing relationship between the two images. The workflow also uses ImageScaleToTotalPixels to normalize both input images before editing. This helps keep the image references aligned and prevents unstable size differences from damaging the result. GetImageSize and EmptySD3LatentImage are used to match the latent canvas to the processed image dimensions, so the output remains consistent with the target image structure. A key part of this workflow is reference control. FluxKontextMultiReferenceLatentMethod is used with the index_timestep_zero method, helping the model understand how to use the visual references during generation. CFGNorm and ModelSamplingAuraFlow help stabilize the FireRed edit model, while KSampler runs an 8-step, low-CFG editing pass with Euler sampling and a simple scheduler. This setup is fast, direct, and suitable for repeated testing.

Qwen 165 downloads
查看公开资料
Anima Acceleration Comparison Testing Workflow
Workflows 2026-05-10

Anima Acceleration Comparison Testing Workflow

This workflow is designed for Anima acceleration comparison testing. Its main purpose is to compare different Anima generation routes, sampling stages, and acceleration settings under the same prompt, same latent size, and similar visual conditions. Instead of only showing one final image, this workflow is built to help creators understand how different sampling strategies affect image quality, speed, structure, detail, and final style. The workflow uses anima-preview.safetensors as the main Anima model, qwen_3_06b_base.safetensors as the text encoder, and qwen_image_vae.safetensors as the VAE. The latent canvas is created with EmptyLatentImagePresets at 1152 x 896, which gives a wide illustration format suitable for anime comparison tests, character previews, and visual benchmark examples. The prompt describes an anime-style fox girl in a snowy night scene holding a sign, which is a useful test prompt because it includes character identity, clothing, environment, lighting, atmosphere, and readable text. The negative prompt is also shared across the comparison routes, suppressing problems such as worst quality, low quality, blurry output, JPEG artifacts, signatures, and artist names. This keeps the comparison cleaner, because each route is tested against the same quality-control conditions. The key value of this workflow is that it does not test Anima as a single black-box output. It breaks generation into multiple sampling paths. Some routes use Anima Preview directly, while other routes apply an additional Anima RDBT-style LoRA path. The workflow includes multiple KSamplerAdvanced nodes with different start and end steps, making it possible to observe how early-stage, mid-stage, and later-stage sampling changes the final image. This is useful for understanding which part of the denoising process controls composition, which part affects details, and which route produces better speed-quality balance. The workflow also uses image concatenation nodes to place outputs side by side. This makes it easier to visually compare results instead of checking images one by one. You can compare whether a route gives better line quality, cleaner character design, stronger lighting, better color harmony, more stable text, or fewer artifacts. The workflow can also export comparison results as animated WEBP, making it useful for tutorial demonstrations, Civitai previews,

Anima 73 downloads
查看公开资料