DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

Anima for Beginners: 16 Workflows, an A/B Gain Test, and the ControlNet V2 Path

An anime base is only useful if a beginner can get a first good image out of it without understanding the whole graph. This release bundles three graphs for three skill levels: a 16-workflow starter pack, an A/B comparison rig for gain nodes, and a ControlNet V2 control path.

A detail that matters more than it looks

The text encoder is loaded as:

CLIPLoader → qwen_3_06b_base.safetensors, type qwen_image

That type string is different from the stable_diffusion type used in other Anima builds. The type tells ComfyUI how to project the encoder output into the model's conditioning space. Getting it wrong does not raise an error — it produces images that look vaguely right but ignore half the prompt.

If your Anima output "feels like it is only reading one word", check the loader type before touching anything else.

Graph 1: 16 workflows in one file (250 nodes)

The starter pack is a single 250-node file rather than 16 files. That is a deliberate structure: every variant shares the same loader trio (anima_baseV10, qwen_3_06b_base, qwen_image_vae), so one file means one place to change the base, and 16 branches to compare.

The LoraLoader in this graph is set to ['None', 0, 0] — a bypassed slot. That is a good default for a starter pack: an empty LoRA slot is easier to fill than a missing node is to add.

Output prefix KSK_Anima_06_Base confirms this graph is the baseline branch produced the comparison image.

Graph 2: the A/B gain rig (154 nodes)

This one exists to answer a single question: does the turbo LoRA actually help?

LoraLoader → anima-turbo-lora-v0.1.safetensors at weight 1.0, with output prefix KSK_D01_A_BASE.

154 nodes for an A/B test sounds excessive until you consider what a fair test needs: the same seed, the same latent, both branches sampled with identical settings, and both outputs saved. Cutting corners on any of those produces a comparison that means nothing.

At weight 1.0 the turbo LoRA is fully active. The honest way to read this graph's output is not "which is better" but "where is the difference visible" — with anime bases it is usually line weight and hair strands before it is composition.

Graph 3: ControlNet V2 (23 nodes)

The smallest graph and the most production-useful:

  • UNETLoader → anima_baseV10.safetensors
  • LoraLoaderModelOnly → anima-highres-aesthetic-boost.safetensors at 1.0
  • EmptyLatentImage at 512×512 — noticeably smaller than the 1024 base used elsewhere
  • Output prefix KSK_Anima_LLite_CN_Final

The 512 base is the right call for a control workflow: the control signal is doing the composition work, so spending latent budget on a larger base buys nothing. The aesthetic-boost LoRA then lifts the result.

Generating at 512 and letting a control network drive structure is a different trade from generating at 1024 and hoping the prompt holds. The first is repeatable; the second is not.

Where this release is weak

  1. 16 branches in one file is hard to navigate. Good for comparison, bad for learning. A beginner opening 250 nodes sees noise, not structure.
  2. The starter pack bypasses the LoRA slot. That is the right default but it means the first thing a beginner will do — load an artist LoRA — is the thing with no example.
  3. The A/B rig does not tell you what changed. It saves two images; reading the difference is still on you.

Practical order for a beginner

  1. Open the control V2 graph first (23 nodes, 512 base). Get one image out.
  2. Move to the 16-workflow file and change only the prompt.
  3. Only then load the turbo LoRA and compare — by that point you know what "normal" looks like.

Workflows

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

  • Anima 16 common workflows collection
  • Anima gain node A/B test
  • Anima ControlNet V2 control workflow

Files

The ComfyUI workflow exports are in the companion repository:

https://huggingface.co/AI-KSK/ksk-video-245-comfyui-workflows

Video walkthrough (Chinese): https://youtu.be/LkB-58hwSlQ
Bilibili: https://www.bilibili.com/video/BV1jALa6xEqH/

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)