Anima Preview 3 in ComfyUI: four graphs, one prompt system, and where the model actually breaks

Anima Preview 3 is the 2026-04 anime-specialised model update from the Anima Preview line. The author walks through the four graphs that matter in daily use — text-to-image, image-to-image, partial repaint and tiled upscale — and spends most of the runtime on the part that actually decides output quality: the prompt system. Everything below is taken from the episode subtitles and from the four workflow JSON files shipped with the episode.

Evidence note: the numbers in the "Parameter basis" section are read from the shipped workflow JSON (anima-preview3-base, four graphs). Claims about quality, resolution limits and sampler behaviour are the author's own reported observations in the video, not an independent re-test.

Where this workflow fits

Use it when you want anime/illustration generation in ComfyUI with a small model stack and a prompt style that is closer to Danbooru tagging than to natural-language prompting. It is a preview model, so it is the right tool for experimentation and style hunting, not for a production SLA.

What Preview 3 actually changed

• Preview 3 is a base-model update only. The VAE and the text encoder are unchanged from Preview 2; you replace the diffusion model and keep the rest of the stack.
• It was trained longer at 1024 resolution and the training set was expanded so that rarer artists are better represented.
• The author's read: prompt adherence and overall image quality are the strong points, and the model is good enough to be used as a fine-tuning base.
• The author still calls the partial-repaint graph "forward-looking": it runs, but a final, fully reliable result is not guaranteed yet.

The four graphs

All four were checked with a single GET: HTTP 200, and the page titles match the graph names above.

Step by step

1. Load the stack. UNETLoader → anima-preview3-base.safetensors (default weight dtype), CLIPLoader → qwen_3_06b_base.safetensors with type stable_diffusion, VAELoader → qwen_image_vae.safetensors. This is the same triple in all four graphs.
2. Use the fixed prompt starter. Preview 3 ships a recommended opening block for positive and negative prompts. Write your actual subject after the starter; for the plain text-to-image case the author simply deletes the one girl fragment from the starter and writes his own subject.
3. Pick a resolution you can afford. The text-to-image graph defaults to the SDXLEmptyLatentSizePicker+ preset 896x1088 (0.82) — about 0.82 MP. The author recommends staying under roughly 1.5 MP, and describes 1.24:1.24 style aspect ratios as "very good".
4. Generate, then decide whether you need a second pass. Two options exist: a pixel-space upscale or the tiled upscale graph.
5. For image-to-image, add a LoadImage, scale with image_scale_pixel_v2, and lower the denoise. The author notes denoise below roughly 0.5 is not very responsive on this preview.
6. For partial repaint, load the image, open the mask editor, paint the region, set the size constraint and run. The graph also contains an RH_LLMAPI_Pro_Node branch that writes the prompt automatically from a system prompt — edit the user half of that prompt to change the request (the author's demo turns a head into a pineapple).
7. For tiled upscale, load the image, set the target size in megapixels, and optionally prepend an artist tag through a string-concat node so the style stays stable.

Parameter basis (from the shipped JSON)

Text-to-image (10 nodes): ClownsharKSampler_Beta with eta 0.35, sampler linear/euler, scheduler beta57, 30 steps, denoise 1, CFG 3; latent preset 896x1088 (0.82).

Image-to-image (13 nodes): eta 0.35, linear/euler, beta57, 30 steps, denoise 0.73, CFG 3; image_scale_pixel_v2 factor 1.0293, 64-pixel multiple.

Partial repaint (16 nodes): eta 0.5, linear/euler, beta57, 30 steps, denoise 1, CFG 3; built from SetLatentNoiseMask + DifferentialDiffusion + GrowMaskWithBlur.

Tiled upscale (52 nodes): eta 0.4, linear/euler, beta57, 25 steps, denoise 0.3, CFG 2; UpscaleModelLoader → 4x_NMKD-Siax_200k.pth; ImageScaleToTotalPixels lanczos at 4 MP; tiling via TTP_Image_Tile_Batch / TTP_Image_Assy; WD14Tagger|pysssss is present for tag reverse-engineering.

Those tiled-upscale numbers (25 steps, CFG 2, denoise 0.3) match what the author says in the video, which is a useful cross-check that the shipped JSON is the graph he was running. The tiled graph also still carries a leftover LoraLoader for Gundam_RX78_Flux.safetensors at 0.85/0.85 — treat it as a demo artefact and bypass it unless you want that gunpla look.

Sampler and scheduler notes

• The scheduler beta57 is supported here. The author switches to it when he wants a more 2.5D / semi-real look.
• The sampler is a non-default one from a plugin rather than a stock ComfyUI sampler. The stated reason: a Reddit user found that this sampler, with the ETA strength kept between roughly 0.2 and 0.5, stabilises the image — most visibly on hands. The author reports that hand collapse becomes much less frequent at some fixed resolutions after switching.
• The JSON's eta values land inside that window (0.35–0.5), which is consistent with the video.

Limits

• Preview 3 is a preview: the repaint graph is explicitly "not a finished result".
• Around 2 MP the preview starts to fall apart. The author recommends keeping generation under about 1.5 MP, and says 1536-level resolution is only officially scheduled for the final stage.
• Higher resolutions such as 1536x1536 look under-trained.
• Hand quality is partly a training problem, not only a sampler problem.
• The author says he will publish a prompt-writing document on Bilibili; that document is not part of this package.

Episode resources

• Workflows (RunningHub, invite code rh-v1111): four posts linked in the table above.
• YouTube: https://www.youtube.com/watch?v=L1Y2EPuuhoU
• Bilibili: https://www.bilibili.com/video/BV1EpDxBcEz4
• No Compshare mirror link was recorded for this episode (优云智算链接：无).