IMPORTANT. Read this first.
This checkpoint is not built for Qwen3.5-4B and not for the expanded adapter.
52-block DiT, but text conditioning is native qwen_3_06b_base only.
Insert blocks after the frozen 28-stem were prepared for that CLIP. Qwen3.5 will look like stock Pro52, not this mix.
Do not use qwen35_4b or Anima-3.8B-expanded_adapter.
Delta Mix
Checkpoint merge on Anima-3.8B (52-block). Anime and illustration first, with extra room for kemonomimi / light furry. Use it like a normal Anima model.
Text encoder: qwen_3_06b (native Anima CLIP). Do not swap in Qwen3.5-4B or an expanded adapter.
VAE: official Anima / Qwen-Image VAE.
Prompt
Positive starter: anime, masterpiece, best quality, ultra detailed, absurdres, sharp focus, intricate details, cinematic dramatic lighting, strong chiaroscuro,
Then the scene in tags or natural language, Anima-style. Rating tag if you use one (safe / sensitive / explicit).
Recommended
Steps: 30
CFG: 4.0–5.5
Sampler: Euler ancestral
Scheduler: normal
Resolution: same as Anima-3.8B (1024-class and up, as VRAM allows).
LoRA
28-block and 2.9B LoRAs need a remap on this 52-block file. Use:
Anima 52 LoRA Patch
https://github.com/bestluner-create/ComfyUI-Anima52-LoRAPatch
Start around 0.65. Do not stack stock Load LoRA on the same old LoRA.
Credits
Anima Base / Anima-3.8B — original authors
Merge references: Aura Rising, Animurr
License follows Anima (non-commercial). Do not sell the weights.
Description
FAQ
Comments (13)
Any way to run on webui forge?
Regular Forge - no.
Forge Neo - probably yes, same as a normal Anima 3.8 checkpoint. I haven’t run this file there myself.
Put the checkpoint in Stable-diffusion, CLIP qwen_3_06b_base, VAE qwen_image_vae. Don’t install the Qwen 3.5 extension, this model doesn’t use it.
Old 28-block LoRAs can land wrong in Forge. If the style breaks, use Comfy.
If your Neo already runs the official Anima 3.8 checkpoint with only qwen_3_06b, this file should load the same way. Old Forge and Neo builds that only know 28-block Anima won’t see 52 layers.
LoRAs made for base Anima may or may not line up. I don’t have a Forge remap. If a LoRA looks off, drop it or use Comfy.
This is probably the most vibrant model I have ever used! the colors just pop! Please keep on working on this model as i see some HELLUVA potential in this!
Thanks. Glad the color landed. This is v1, I’m still working on it.
Second question: Right now is there a way to do negative loras like Bad_Anatomy?
Workflow: https://drive.google.com/file/d/1dDBN4uHxacQnBTg9bybqVuWoOaF5RDdJ/view?usp=sharing
Yeah you can push a LoRA negative, same idea as Bad Anatomy.
In Forge Neo put it in the positive prompt with a minus weight, not in the negative box:
<lora:YourLoraName:-0.4>
Forge often ignores LoRA tags if you drop them in the negative prompt.
One catch: the usual SDXL / Pony Bad Anatomy LoRAs won’t do anything useful on this model. Anima is a different architecture, those keys don’t land. You need an Anima-trained LoRA, then run it at about −0.3 to −0.5 and see.
If you just want fewer broken hands, the short negative is enough:
worst quality, low quality, score_1, score_2, score_3, bad hands, extra fingers, fewer digits, missing fingers
The giant Illustrious “ugly, bad anatomy, long body…” list is mostly leftover from XL and tends to fight Anima more than it helps.
@bestluner180 You are so helpful thank you!!
bf16 is faster than fp16 even if you transform both into int8convfort
just saying. It will be 25% speed difference.
Correction: Only when loras are involved.
Correction2: Only when you use incompatible lora loader.
Thanks - useful if LoRAs are in the graph.
Which loader did you use?
This file expects Anima 52 LoRA Patch
(28-block → stem only, 40-block → stem+Gen-1).
Stock Load LoRA / other 52 remaps will skew the timing.
If you have a screenshot of the node + bf16 vs fp16, I can check.
@bestluner180 (edit: i was wrong)*
Used this node for convertion.
Loaded with Diffusion Model Loader standard one with default weight.
I use this node for 3.9B lora loading.
*forget that i said anything, the time difference happens only when you use regular lora loader (incompatible one)
When used 3.9B lora loader the time difference was 1% in favor of BF16 converted to Int8Conv so in margin of error.
So this speed diff was only a curious thing that maters not.
@AltairTheArc Thanks for the correction.
That matches what I’d expect: stock LoRA loader on a 52-block file
is the slow/wrong path. Bridge / 52 LoRA Patch should be within noise.
bf16 and fp16 here are the same mix - pick what your loader wants.
best model !!