Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 3 min read

WAI-ANIMA: A Flux LoRA in an Anime Model, and the Encoder Type That Trips Everyone

WAI-ANIMA is a next-generation anime illustration base built on the Anima lineage. This release covers style workflow, style retention, size limits, sampler choice and hires fix β€” and along the way it exposes two configu

WAI-ANIMA is a next-generation anime illustration base built on the Anima lineage. This release covers style workflow, style retention, size limits, sampler choice and hires fix β€” and along the way it exposes two configuration traps worth understanding before you run it.

The base stack

All three graphs load:

Node File
UNETLoader waiANIMA_v10(anima3οΌ‰.safetensors
CLIPLoader qwen_3_06b_base.safetensors
VAELoader qwen_image_vae.safetensors

The filename contains full-width parentheses β€” (anima3οΌ‰. That is not a transcription error, it is the actual filename. On Windows this loads fine. On a Linux box or inside a container using a non-UTF-8 filename pipeline, this fails silently: ComfyUI reports the model as missing, and the usual reaction is to assume the download is corrupted. Check the filename encoding before re-downloading.

Trap 1: the CLIP loader type is not consistent

Workflow Nodes CLIP type
Text-to-image, latent upscale + hires fix 16 qwen_image
Block/tiled upscale 26 qwen_image
Image-to-image 13 stable_diffusion

Same encoder file, same base model, two different type strings. The type changes how the encoder output is projected into conditioning space. Wrong type does not error β€” it produces an image that responds weakly to the prompt.

This is the same inconsistency seen across the broader Anima family. The rule that saves time: copy the loader node from the workflow you are extending; never retype it.

Trap 2: a Flux LoRA inside an Anima graph

The block-upscale workflow loads:

LoraLoader β†’ Gundam_RX78_Flux.safetensors at weight 0.85

The filename says Flux. The base is Anima. These are different architectures, and a Flux LoRA is not architecturally valid for an Anima base.

What actually happens depends on how the LoRA tensors align with the target model. In many cases it loads without error and applies some transformation β€” which is worse than failing, because the output looks intentionally styled and you have no way to know it is a wrong-architecture artifact rather than the LoRA's intended effect.

If you are using this workflow as a template, clear that LoRA slot first. It is almost certainly left over from the author's testing rather than being a recommendation.

The 0.85 weight is also suspicious in the same way β€” a non-unity weight on an incompatible LoRA is not a meaningful tuning choice.

Workflow 1: text-to-image with latent upscale and hires fix (16 nodes)

Base latent 512Γ—512, then latent-space upscale, then hires fix. Two resolution stages, both in latent space, which is why 16 nodes is enough.

512 base is the right starting point for anime: the style is carried by line work and flat colour, both of which survive a smaller base better than photographic detail does.

Workflow 2: image-to-image (13 nodes)

The smallest graph and the CLIP-type outlier. This is the "I have a rough, make it match WAI-ANIMA's style" workflow β€” useful for bringing an existing sketch or a generated image into a consistent look.

Workflow 3: block upscale (26 nodes)

Tiled upscaling for large output. Tiling is kinder to anime line art than a single resample because each tile retains its own frequency content, so line weights do not average out across the image.

Where this release is weak

  1. Two real traps (encoder type, foreign-architecture LoRA) that both fail silently.
  2. 13-node image-to-image has essentially no validation. A wrong loader type or a bad LoRA produces plausible output.
  3. No sampler documentation in the files. The episode covers sampler choice, but the workflows do not carry that reasoning β€” you have to get it from the video.
  4. Full-width filenames. A portability landmine.

Practical order

Open the 16-node text-to-image graph, clear any LoRA slots that do not belong to the Anima family, confirm the CLIP type is qwen_image, and generate one image. Only after that baseline works should you touch the upscale or image-to-image graphs.

Workflows

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

  • WAI ANIMA block/tiled upscale
  • WAI ANIMA image-to-image
  • WAI ANIMA text-to-image with latent upscale and hires fix

Files

https://huggingface.co/AI-KSK/ksk-video-225-comfyui-workflows

Notes

  • Workflow files only. Model weights are not included; each keeps its own licence.
  • Tutorial exports, not a product.
  • Node counts, CLIP types and LoRA values quoted above are read from the published JSON files.
  • Independent AI-KSK release. Not affiliated with the model authors or RunningHub.
πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.