DEV Community

AI-KSK
AI-KSK

Posted on

Anima Base 1.0: Text, Image, Tiled Upscale and Inpaint Workflows With LLLite Control

#comfyui #ai #tutorial

Anima Base 1.0 is the anime-illustration base model the author has tracked through its preview releases. This episode ships four graphs around it: text-to-image with control and face/hand repair, image-to-image with the same repair tail, tiled upscaling, and masked repaint. The author reports in the video that 1.0 lifts usable resolution to roughly 2.3 megapixels and that the LLLite control path finally behaves after the 2000-step training update.

Evidence note: four workflow JSONs were shipped with the episode, so node names, model files and sampler settings below are read from those files; behaviour claims are the author's, as reported in the video.

When to use it

  • You want an anime/illustration base you can drive with tags and natural language and also reference artist styles with @artist.
  • You need a control (depth/LLLite) pass plus a face/hand repair tail in one graph.
  • You have an image that needs a large tiled upscale (author's framing: four tiles, ~8 MP).
  • You need a repaint graph for control-bearing edits only; the author is explicit that this is the weakest of the four.

What is different in this workflow

  • Control is a first-class branch: AnimaLLLiteApply + AIO_Preprocessor sit before the sampler, so you can bypass the control group (Ctrl+B) instead of deleting nodes.
  • Two samplers, one prompt: a base pass and a high-resolution pass, each with its own ClownsharKSampler_Beta, then LatentUpscaleBy and two FaceDetailer passes.
  • The LLM prompt node is optional: RH_LLMAPI_Pro_Node pre-formats prompts into Anima's preferred shape; delete it and type prompts by hand.
  • Tiled upscale is a separate graph, not a chain inside the text-to-image graph.

Step by step

  1. Load the text-to-image graph. Keep RH_LLMAPI_Pro_Node if you want prompt formatting, otherwise delete it and fill the positive/negative encoders yourself.
  2. Set size. The author reports ~2.35 MP is now safe and runs the resolution control at 2.3.
  3. Keep the control group enabled if you are feeding a reference image; Ctrl+B to bypass it when not.
  4. Optional Turbo route: load anima-turbo-lora-v0.1, drop to 8-12 steps, CFG 1, and enable NAG.
  5. Run; the latent upscale plus the two FaceDetailer passes run in the same queue.
  6. Image-to-image: load the reference into LoadImage, encode, run with the same repair tail. The author uses this as style reference, not as an edit model.
  7. Tiled upscale: load the image, set the target size, keep four tiles; the author reports the ceiling around 9.4 (8 MP+ in total).
  8. Inpaint: paint the mask, GrowMaskWithBlur + SetLatentNoiseMask, run.

Parameter basis

  • Common to all four graphs: UNETLoader -> anima_baseV10.safetensors; CLIPLoader -> qwen_3_06b_base.safetensors (type qwen_image); VAELoader -> qwen_image_vae.safetensors.
  • Text-to-image (37 nodes): LoraLoaderModelOnly -> anima-highres-aesthetic-boost.safetensors weight 1.0; two ClownsharKSampler_Beta at 30 steps, linear/euler, beta57, raw widget tuples [0.35, 'linear/euler', 'beta57', 30, -1, 0.55, 2.88, seed, 'randomize', 'standard', true] (base pass) and [0.35, 'linear/euler', 'beta57', 30, -1, 1, 3.01, seed, 'randomize', 'standard', true] (high-res pass); LatentUpscaleBy; 2x FaceDetailer with 3 UltralyticsDetectorProvider; NAGuidance.
  • Image-to-image (33 nodes): same CLIP/VAE, but the second UNETLoader carries waiANIMA_v10（anima3）.safetensors and the CLIP type is stable_diffusion; sampler 30 steps, beta57, CFG field 3.0.
  • Tiled upscale (27 nodes): LoraLoader -> anima-turbo-lora-v0.1.safetensors at 0.85/0.85; sampler 20 steps, CFG field 2.0; TTP_Image_Tile_Batch / TTP_Image_Assy with TTP_Tile_image_size; WD14Tagger|pysssss for the style text; ImageUpscaleWithModel.
  • Inpaint (15 nodes): DifferentialDiffusion + SetLatentNoiseMask + GrowMaskWithBlur; sampler 30 steps, CFG field 3.0, denoise 1.0.

Limits

  • The author calls the LLLite control path "usable" only after the 2000-step update, and ranks local repaint as still weak — other editing models do it better.
  • Turbo LoRA is a trade: the author says it degrades style and detail, and prefers 30-50 steps when the hardware allows.
  • Latent upscale "looks neither broken nor sharp"; the author has not solved that softness.
  • The tiled upscale keeps style only if you supply style text (the graph exposes a text input for it).

Episode resources

Top comments (0)