Anima 2.9B in ComfyUI: Six Workflows, From Official T2I to Tiled Redraw and Smart Img2Img

Anima went from 2B to 2.9B, and the interesting part for ComfyUI users is not the parameter count — it is that the same checkpoint now holds up across six different jobs. This episode ships all six graphs: the official text-to-image baseline, a 1.5x upscale pass, a three-stage render, a tiled high-resolution redraw, a text-to-image route with face and hand repair, and a Qwen3-VL-captioned image-to-image route.

No inference was run for this release. Everything below is read from the saved nodes: sampler values, latent sizes, denoise settings and the detector/upscaler files. There is no timing, no VRAM number and no quality comparison.

The shared base

All six load Anima-2.9B-preview-v1.safetensors through a UNETLoader, encode text with qwen_3_06b_base.safetensors (type stable_diffusion) and decode through qwen_image_vae.safetensors. The official negative block — the long "worst quality, lowres, old, ancient..." string — is saved in each graph's negative encoder.

Where the files differ is the sampling recipe. The official baseline values used through the set are 28 steps, CFG 5, res_multistep and linear_quadratic; each workflow changes the denoise and the number of passes around those values.

One trap worth naming: the official T2I file's active sampling chain is not the rgthree "KSampler Config" node. The config node stores 28 steps / CFG 5 as a reference, but the live path is SamplerCustomAdvanced with KSamplerSelect euler, BasicScheduler sgm_uniform at 50 steps denoise 1 and CFGGuider at CFG 8. If you quote the config node as the setup, you are quoting the wrong thing — I did exactly that check before writing this.

01 — The official baseline

Two EmptyLatentImage canvases (1216x832 and 1152x1536) switch through an ImpactSwitch, so one file gives you landscape and portrait without rewiring. Output prefix is Anima_Sample/preview_2.9b. The prompt is a demo scene with two Genshin characters; treat it as a template, not content.

02 — 1.5x upscale: the simplest tuning surface

Generate at 1024x1024, upscale the latent 1.5x with nearest-exact, then run the second KSampler at denoise 0.35. The upscale happens on the latent, not on the decoded image, so the second pass sees the enlarged latent directly. If the result looks soft, raise the 0.35 slightly; if the subject starts changing, lower it.

03 — Three-stage render: same idea, two steps further

The base pass runs at denoise 1, then two LatentUpscaleBy nodes at 1.25x each carry the image through denoise 0.35 and 0.3. Final size is the product of those factors, so if you edit either 1.25, re-check the output resolution. This is the file where "more stages" costs the most time for the least visible change if your base prompt is already sharp — try 02 first.

04 — Tiled redraw: the memory-bounded high-res route

This one takes an existing image instead of generating from noise: 4x_NMKD-Siax_200k upscale, then ImageScaleToTotalPixels to 4 MP, then a 2x2 tiling with 0.35 overlap, VAE encode, a KSampler at denoise 0.20, and tiled VAE decode. The WD14 tagger branch builds the caption from the input image.

The 0.20 denoise is the whole point: it re-renders detail without repainting the composition. Raise it and tiles start disagreeing at the seams. The overlap is what hides those seams, so if you change the tile count, keep the overlap and look at the joins.

05 — Face and hand repair: two detailers, three detectors

Base pass at 512x512, FluxResolutionNode to 1.0 MP, denoise 1, 1.5x latent upscale, then denoise 0.35 — and then the interesting part: two FaceDetailer nodes fed by three UltralyticsDetectorProvider models (bbox/face_yolov8n.pt, bbox/hand_yolov8s.pt, and a segmentation model segm/PitHandDetailer-v2-Test-v9c.pt) plus a SAMLoader with sam_vit_b_01ec64.pth. Detailer denoise is 0.30 / 0.35.

The failure mode is silent: if the detector does not find a hand, no repair happens and nothing warns you. That is why the graph saves two images — base_preview and face_hand_refined — so you can compare.

06 — Smart img2img: let Qwen3-VL write the prompt

Load an illustration, scale it to about 1.05 MP, VAE-encode it, and let a Qwen3VLProcessor node produce an Anima-compatible prompt (quality tags, year tags, character count, then subject and scene). The KSampler runs at denoise 0.7.

At 0.7 the caption matters more than the input image. Read the generated prompt before queueing; if it misreads the character or the outfit, the output will be confidently wrong. Lower the denoise if you need the composition preserved harder.

How I would sequence a first run

Start with 01 or 02 to see what the checkpoint does at 1 MP. Only go to 03 if the two-pass result is visibly soft. Use 04 when you already have a good image and want a bigger version; use 05 when anatomy is the problem; use 06 when you have a reference illustration and want it restyled.

What this is not

Not a benchmark, and not a license grant. The video's framing about 2.9B, character reproduction and "80 test images" is the author's description. The files carry no grant for the checkpoint, the encoder, the VAE, the upscaler, the detectors or the demo art — check those separately.

Getting the workflows

Published by this channel. The graphs are free; the checkpoint, VAE, encoder, upscalers, detectors, taggers and example media are not bundled.

• Hugging Face: AI-KSK/anima-29b-six-workflows-295 — six JSONs, source covers, SHA256 manifest. Start here.
• ModelScope (魔搭): Sunjian520/anima-29b-six-workflows-295
• Civitai: 01 official T2I · 02 1.5x upscale · 03 three-stage · 04 tiled redraw · 05 face/hand repair · 06 smart img2img
• Run without a local GPU: 01 · 02 · 03 · 04 · 05 · 06 — these carry the author's invite code rh-v1111.
• Tutorial: YouTube · Bilibili

Verify workflow, node, checkpoint, encoder, VAE, upscaler, detector, tagger, VLM, input-image and cover rights separately. No blanket commercial or redistribution license is asserted for the graphs or the example media.