DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

Anima Base 1.0: Four Workflows, and a Text-Encoder Type That Changes Between Them

Anima Base 1.0 is an open anime illustration base, and this release ships four workflows that cover the usual production needs: text-to-image with control and face/hand repair, image-to-image with the same, tiled upscaling, and inpainting.

The workflows are worth studying together because they expose a configuration difference that will silently degrade your output if you miss it.

The stack

All four load the same three:

Node File
UNETLoader anima_baseV10.safetensors
CLIPLoader qwen_3_06b_base.safetensors
VAELoader qwen_image_vae.safetensors

Verified same base, same encoder, same VAE. What differs is the CLIP loader type.

The type inconsistency — read this before you edit anything

Checking the actual widget values in the four published files:

Workflow Nodes CLIP type
Text-to-image + CN + face/hand repair 37 qwen_image
Block/tiled upscale 27 qwen_image
Image-to-image + CN + face/hand repair 33 stable_diffusion
Inpainting 15 stable_diffusion

Two workflows say qwen_image, two say stable_diffusion, for the same encoder file against the same base model.

This matters because the type string changes how ComfyUI projects the encoder output into the model's conditioning space. Get it wrong and nothing errors — the image is generated, it just responds to your prompt more weakly than it should. It reads like "the model is ignoring part of what I wrote", which is exactly the symptom people misdiagnose as a bad prompt.

Practical rule: copy the loader node from the same workflow you are extending, do not retype it.

Workflow 1: text-to-image with control and repair (37 nodes)

The most complete of the four. EmptyLatentImage at 512×512 — small on purpose. The control network does the composition work, so latent budget goes into the control path rather than the base resolution.

LoraLoaderModelOnly → anima-highres-aesthetic-boost.safetensors at weight 1.0 then lifts the result. Building at 512 and boosting is more repeatable than building at 1024 and hoping.

The face and hand repair stages are separate — which is the right structure. Hands and faces fail for different reasons, and a single "detail fix" pass tends to improve one at the expense of the other.

Workflow 2: tiled upscale (27 nodes)

LoraLoader → anima-turbo-lora-v0.1.safetensors at weight 0.85, not 1.0.

The 0.85 is interesting. Turbo LoRAs at full weight can produce a characteristic flatness on anime line art; backing off to 0.85 keeps most of the speed and restores some of the line variation. This is one of the few places in these graphs where a fractional weight is clearly a considered choice rather than a leftover.

Tiled upscale rather than a single resize pass means the graph can handle a large output without one giant allocation — and for anime specifically, tiling is kinder to line art than an implicit resample because each tile keeps its own frequency content.

Workflow 3: image-to-image with control and repair (33 nodes)

Same structure as the text-to-image graph but takes an image instead of an empty latent, keeping the CN and repair stages. This is the workflow for "I have a rough, make it good" — sketch to illustration, low-res to clean.

Note the CLIP type here is stable_diffusion, unlike its text-to-image sibling.

Workflow 4: inpainting (15 nodes)

The smallest graph, and the one you will use most once you are iterating: 15 nodes means it is close to the minimum for a working inpaint. Same base, same VAE, stable_diffusion encoder type.

For anime work inpainting is disproportionately useful — fixing a hand, replacing a prop, cleaning a background — because the surrounding line art constrains the fill tightly.

Where this release is weak

  1. The CLIP type inconsistency is a real trap. It is not documented anywhere in the files and produces a soft failure.
  2. Two encoding conventions in one suite means an image produced by the text-to-image graph and one from the image-to-image graph are not directly comparable — a difference you observe may be the encoder projection, not the workflow.
  3. Face and hand repair are inside the main graphs, not separable. If hand repair is what you need, you run the whole 37-node graph.
  4. 512 base. Convenient for control work, but it means the aesthetic-boost LoRA is doing real structural work, not just polishing.

Practical order

Start with the 15-node inpaint to confirm your models load and your prompt style works. Then the 37-node text-to-image. Leave the upscale and image-to-image graphs until you have a base you actually want to keep.

Workflows

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

  • Anima Base text-to-image + CN + face/hand repair
  • Anima Base image-to-image + CN + face/hand repair
  • Anima Base block/tiled upscale
  • Anima Base inpainting

Files

The ComfyUI workflow exports are in the companion repository:

https://huggingface.co/AI-KSK/ksk-video-241-comfyui-workflows

Video walkthrough (Chinese): https://youtu.be/FM5dGuTnz8o
Bilibili: https://www.bilibili.com/video/BV1ci5q65EGQ/

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 CLIP type 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)