DEV Community

AI-KSK
AI-KSK

Posted on

Anima Preview 2 in ComfyUI: four graphs, one enhancement node, and a conflict worth flagging

#comfyui #ai #tutorial

Anima Preview 2 is the anime-specialised preview that added character enhancement and better natural-language understanding over Preview 1, and it was designed to be fine-tuned. This episode walks four graphs — text-to-image, image-to-image, tiled upscale, and a prompt-line batch style test — and then spends real time on the community Anima-Enhancer node, whose parameters are unusual enough to deserve a table.

Evidence note: all four graphs ship as JSON and the parameters below are parsed from them. Where the video's spoken guidance conflicts with the shipped defaults, both are reported rather than smoothed over — see "A conflict worth flagging". The Anima-Enhancer parameter walkthrough, the 1-megapixel training caveat and the tile-seam advice are the author's statements in 字幕.srt.

Where this fits

Anime and illustration work in ComfyUI with a compact stack: anima-preview2 on a Qwen3 text encoder and the Qwen-Image VAE. It is good at style work and cheap enough to iterate. It is also, emphatically, a 1024-class model.

The four graphs

Graph Nodes RunningHub
Anima Preview2 文生图 14 https://www.runninghub.ai/post/2033542967683715074/?inviteCode=rh-v1111
Anima Preview2 图生图 13 https://www.runninghub.ai/post/2033542974574956546/?inviteCode=rh-v1111
Anima Preview2 分块放大 25 https://www.runninghub.ai/post/2033542982825152514/?inviteCode=rh-v1111
PromptLine + Anima Preview2 艺术家批量风格测试 16 https://www.runninghub.ai/post/2033542990416846849/?inviteCode=rh-v1111

All four fetched: HTTP 200, titles match.

Shared stack in every graph: anima-preview2.safetensors (UNET), qwen_3_06b_base.safetensors (CLIP, type stable_diffusion), qwen_image_vae.safetensors (VAE).

Parameter basis (from the JSON)

Text-to-image (14 nodes): two KSampler nodes — 20 steps, CFG 2.5, denoise 0.65 and 30 steps, CFG 3, denoise 1; sampler dpmpp_2m_sde_gpu, scheduler sgm_uniform; an image_scale_pixel_v2 at factor 1.6.

Image-to-image (13 nodes): 40 steps, CFG 3, denoise 0.75, dpmpp_2m_sde_gpu / sgm_uniform; scale factor 1.0486.

Tiled upscale (25 nodes): 20 steps, CFG 2, denoise 0.25; UpscaleModelLoader = 4x_NMKD-Siax_200k.pth; ImageScaleToTotalPixels lanczos at 3 MP.

Batch style test (16 nodes): two samplers, 30 steps / CFG 3 / denoise 1 and 20 steps / CFG 2.5 / denoise 0.65 — the same pair as the text-to-image graph, wrapped by the PromptLine node so one prompt sweeps many artists.

A conflict worth flagging

In the video the author says to use 30-50 steps and CFG 4-5, with lower steps pairing with lower CFG. The shipped JSON uses 20-40 steps and CFG 2.5-3. Both are reported here. The JSON is what actually produced his demo images; the spoken range is his general guidance for the model. If you are reproducing his results, start from the JSON values and treat the wider range as headroom.

Step by step

  1. Pick a resolution near one megapixel. His recommendation is 1024x1024 or 896x1152. The reason is specific: the Preview model was only briefly trained at 1024, and beyond that both structure and detail collapse. Within 1024x1024 it holds up.
  2. Write the prompt in the documented order. Artist calls must be prefixed with @; without it the artist's influence drops sharply. The model accepts natural language, tags, or a mix of both — it reads two prompt "systems".
  3. Generate. For a second pass, either upscale in pixel space or use the tiled graph.
  4. Image-to-image is the fast style converter: one light prompt reversal is enough to move a photo into anime style. Real photos work as inputs.
  5. Tiled upscale needs restraint: high denoise across tiles makes the seams obvious. The author uses 20 steps with a reduced CFG, and WD14 tagging for the prompt reversal because the model is tag-native and WD14 is fast.

The Anima-Enhancer node

A community node that improves generation quality and, optionally, speed:

Parameter Value Meaning
block_indices 345 Which key layers get an extra pass. The default is described as the safe setting: hair, cloth folds and line edges come out more solid. Raising it to 8 makes the boost much stronger.
denoise (start) 0.5 Where enhancement kicks in. At 0.5 the model first builds composition itself, then the second half refines detail. Lower = earlier and more aggressive, easily overdone; higher = later, more like a post-process.
end 1 Common value. If the second half looks too heavy or the image feels hard, pull it back to 0.95 or 0.9.
enable_spectrum false Off = pure enhancement. On = enhancement plus speedup.
spectrum value 1 0.2 / 0.22 The most important one. Higher accelerates harder, lower is safer. Above roughly 0.22 the image starts to float and look wrong.
spectrum value 2 8-12 Reads as how complex the "head turn" is during acceleration. Not monotonically better; 8-12 is the stable band.
spectrum value 3 0.5 Higher is more conservative and more stable. Raise it if acceleration misbehaves.
warm-up steps 6 Hold back for the first few steps, watch, then commit. Lower to go faster.

His summary: default parameters are usually enough; turn spectrum on only when you actually want speed.

Limits

  • Preview model: 1024-class training, quality collapses above about one megapixel.
  • Artist tags need the @ prefix or they are largely ignored.
  • Tiled upscale is denoise-sensitive; pushing it produces visible seams.
  • The enhancer's spectrum mode can make images "float" — 0.2-0.22 is the safe band, not a starting point to exceed.
  • The spoken step/CFG guidance and the shipped JSON disagree (30-50 / 4-5 versus 20-40 / 2.5-3); reproduce from the JSON, not the narration.
  • The JSON is a preview-era snapshot; the model line has since moved to Preview 3.

Episode resources

Top comments (0)