DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

Anima UltraReal: Making an Anime Model Render Real People, and Three Ways It Goes Wrong

Anime models and photoreal models are trained on different distributions, and asking one to produce the other usually gives you an anime character with photographic skin texture — uncanny rather than realistic. UltraReal is a fine-tune that attacks that gap directly.

This release ships three graphs: the UltraReal generation path, an 8-step fast path on a different anime base, and a multi-model comparison rig for anime-to-photoreal conversion.

Graph 1: UltraReal generation (38 nodes)

Node Value
UNETLoader ultrarealFineTune_v10AnimaPreview1.safetensors
CLIPLoader qwen_3_06b_base.safetensors, type qwen_image
VAELoader qwen_image_vae.safetensors
EmptyLatentImage 1024×1024

The fine-tune is built on Anima Preview 1, and the name says it — AnimaPreview1. That matters if you are mixing it with later Anima builds: the base it was tuned against is not the current Anima Base.

1024×1024 is the right resolution for photoreal. Unlike the 512-based control workflows, photoreal detail lives in skin pores, hair strands and fabric weave — all of which vanish below about a megapixel.

The face and hand repair in this graph is split, not unified. 38 nodes for a text-to-image graph means the repair stages have real structure behind them.

Graph 2: 8-step fast path (18 nodes)

A different base entirely:

  • UNETLoader → waiANIMA_v10(anima3).safetensors
  • LoraLoaderModelOnly → anima-turbo-lora-v0.1.safetensors at weight 1.0
  • EmptyLatentImage at 512×512

The filename contains full-width parentheses — waiANIMA_v10(anima3). That is a real filename with CJK punctuation, not a transcription artifact. On Windows this is fine; on some Linux toolchains non-ASCII filenames inside ComfyUI model directories cause silent lookup failures. Worth knowing before you move a workflow between machines.

512 base with an 8-step turbo LoRA is the opposite trade from the UltraReal graph: speed over detail, resolution deferred.

Graph 3: the anime-to-photoreal comparison rig (83 nodes)

The interesting one, because it does not use Anima at all for the comparison leg:

Node Value
UNETLoader fluxKleinFP8_flux2KleinBase9bFp8.safetensors, dtype fp8_e4m3fn
CLIPLoader qwen_3_8b_fp8mixed.safetensors, type flux2
VAELoader flux2-vae.safetensors
LoraLoaderModelOnly klein_9B_Turbo_r128.safetensors at 1.0

Note the CLIP is qwen_3_8b here, not the qwen_3_06b used in the Anima graphs. Flux 2 Klein takes a larger Qwen encoder. Two different Anime graphs and one Flux graph means three encoders across three workflows — a maintenance trap if you assume "the text encoder" is one file.

83 nodes for a comparison rig is reasonable: fair comparison needs matched seeds, matched latents, matched step counts, and both outputs saved.

Where the anime-to-photoreal idea breaks

  1. Anime proportions do not become human proportions. UltraReal changes the rendering, not the anatomy. A character with a stylised head-to-body ratio renders as a real person with that ratio — which reads as wrong rather than stylised.
  2. Eyes are the tell. Anime eyes are drawn with a rendering convention that has no photoreal equivalent. UltraReal narrows the gap; it does not close it.
  3. The FP8 quantisation in the comparison leg. fp8_e4m3fn saves memory and costs micro-detail. For a comparison against a full-precision path, that is a confound worth naming.
  4. Three encoders, three conventions. Mixing loader types across these graphs is the same soft-failure trap as elsewhere in the Anima family.

Practical order

Run the 18-node graph first — it is fast enough to establish your prompt style. Move to the 38-node UltraReal graph once the prompt is settled. Use the comparison rig only when you are deciding between model families, not between prompts.

Workflows

Three graphs, published on RunningHub with invite code rh-v1111:

  • Anima UltraReal photoreal anime-character generation with split face/hand repair
  • Anima 8-step fast anime image
  • Anime-to-photoreal multi-model comparison

Files

The ComfyUI workflow exports are in the companion repository:

https://huggingface.co/AI-KSK/ksk-video-235-comfyui-workflows

Video walkthrough (Chinese): https://youtu.be/ (see episode description)
Bilibili: see episode description

Notes

  • Workflow files only. Model weights are not included; each model keeps its own licence.
  • These are tutorial exports, not a packaged product.
  • Node counts and loader values quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with the model authors or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)