Delta Mix 3.8B Qwen 3.5 - v4
A recut of the same line, not a new family. v3 tightens frame stability and edge clarity: lines and object borders stay even instead of swimming. Same dual brief — human and furry.
Samplers: dpmpp_2m / euler_a / res_multistep
Schedulers: res_uniform / normal / beta
Delta Mix 3.8B Qwen 3.5 - v3
Anatomy and hands: fewer collapsed fingers and joint errors on the same test prompts.
Q35 path: the base holds pose / crop more steadily when the 4B encoder is in the graph. Stem (W): replaced. W is now a custom assembled 2B stack, not the previous floor.
Delta Mix 3.8B Qwen 3.5 - v2
Anima 3.8B mix for the official v1.1 chassis and Qwen 3.5 Connector v2.
Source: CircleStone Labs / lylogummy Anima 3.8B v1.1.
This bake fills W + Gen-1 + Gen-2 (Delta Mix overlay).
Connector v2 and the Q35 watch slots stay as in v1.1.
Qwen 3.5 is attached from a sidecar, not baked into this file.
Cleaner image than v1. Fewer W / Gen-2 clashes.
Fine detail and extra fingers still slip — for v3.
Delta Mix 3.8B Qwen 3.5 -v1
Anima 3.8B mix for the official v1.1 chassis and Qwen 3.5 Connector v2.
Source: CircleStone Labs / lylogummy Anima 3.8B v1.1.
The DiT is a Delta Mix overlay (stem + light Gen-1). Gen-2 insert slots and the Connector stay as in v1.1. Qwen 3.5 is attached from a sidecar, not baked into this file.
Downloads
This checkpoint
Sidecar anima_connector_only → same folder (or !Anima3.8/)
Nodes: ComfyUI-Anima-Delta-Q35
Official runtime: comfyui-anima-3-8B (MIT) — Qwen 3.5 4B loader and Connector class
Text encoders from the v1.1 guide: Qwen 0.6B + Qwen 3.5 4B
Lora loader: ComfyUI-Anima-Delta-Q35-LoRA
UNET → DeltaMix Q35 LoRA → Attach → Prompt (id DeltaMixQ35LoraLoader)Graph
UNETLoader this bake (~7.3G)
→ DeltaMix Q35 Attach sidecar (190 tensors) or anima38B_11.safetensors
→ DeltaMix Q35 Prompt CLIP 0.6B + CLIP 4B
expanded → CFG / SamplerPrompt Dual Delta: pos = quality tags + scene, neg = negative tags only.
Use expanded for sampling. native is 0.6B-only.
Extract once: DeltaMix Q35 Extract Connector on anima38B_11.safetensors. Do not extract again unless v1.1 changes.
This file does not include the Connector (~1.2G) or the 4B encoder.
Do not load *Q35bundle in this graph. Bundles are for the official v1.1 loader.
Sampler
Same stack as the v1.1 graph: RandomNoise → CFGGuider → SamplerCustomAdvanced + BasicScheduler. A plain KSampler is fine if the numbers match.
Steps 40. CFG 6. Denoise 1.0.
Sampler: er_sde, Euler, or whatever you already run on 3.8.
Scheduler: normal for this mix. beta if you want the stock v1.1 cadence.
Empty latent 1024×1024. 1920×1080 is stable.
Connector strength 1.0. Both Prompt nodes use the expanded output.
Do not match seeds between the official V2Loader graph and this one. The noise path is different. Compare seeds only inside the same graph.
Credits
V-4
Anima - CircleStone Labs
3.8B v1.1 chassis and Semantic Connector v2 - lylogummy
comfyui-anima-3-8B - GumGum10 (MIT)
Extract / Attach / Prompt - Delta Mix
Method: Anima Delta Mix — 2B into 2.9B / 3.8B
Custom - G1
Custom - G2
DeMix - core
V-3
Anima - CircleStone Labs
3.8B v1.1 chassis and Semantic Connector v2 - lylogummy
comfyui-anima-3-8B - GumGum10 (MIT)
Extract / Attach / Prompt - Delta Mix
Method: Anima Delta Mix — 2B into 2.9B / 3.8B
Aura_Rising - G1
aniMurr - G2
Custom - core
V-2
Anima - CircleStone Labs
3.8B v1.1 chassis and Semantic Connector v2 - lylogummy
comfyui-anima-3-8B - GumGum10 (MIT)
Extract / Attach / Prompt - Delta Mix
Method: Anima Delta Mix — 2B into 2.9B / 3.8B
Aura_Rising - G1
aniMurr - G2
anima_aesthetic - core
V-1
Anima - CircleStone Labs
3.8B v1.1 chassis and Semantic Connector v2 - lylogummy
comfyui-anima-3-8B - GumGum10 (MIT)
Extract / Attach / Prompt - Delta Mix
Method: Anima Delta Mix — 2B into 2.9B / 3.8B
Aura_Rising - G1
Chosen-mix_anima - core
Description
FAQ
Comments (8)
Can I use this with Forge Neo? I'm using version 2.28, but I can't get it to stop producing black images.
Forge Neo 2.28 can run official Anima 3.8B + Qwen 3.5, but this file is not that bundle.
This checkpoint is a ComfyUI Q35 bake: UNET + Semantic Connector sidecar + native Qwen 0.6B together with Qwen 3.5 4B. Black frames almost always mean Forge loaded it as a plain 2B / SDXL-style path — wrong latent, no connector, or the 4B encoder in the wrong slot.
If you still want to try Forge:
Keep qwen_image_vae and qwen_3_06b_base in the top VAE / Text Encoder dropdown. Do not put qwen35_4b there.
Install the 3.8B extension: GumGum10/forge-anima-3.8B (needs Neo 2.28+). The 4B encoder is selected inside that extension, not in the top bar.
Shift around 3–4.5. Empty / 4-channel latent will stay black.
I build and test this page in ComfyUI portable, not Forge. If official Anima 3.8B v1.1 already goes black in your 2.28 install, this merge will too — fix the official graph first. ComfyUI is the supported path for this checkpoint.
@bestluner180 thanks
@bestluner180
369
I managed to solve it; actually, I just had to update Forge Neo to version 2.29 and reinstall the extension. I have to say the jump in quality is huge—I went from 2b to this, and honestly, I doubt I'll ever go back to Anima 2b. The model is truly incredible. Anyway, thanks a lot; your comments led me to try things that helped me find the solution.
@nutakugame007654 Good to hear 2.29 + a clean reinstall of the 3.8B extension fixed it. Glad the jump from 2B landed.
Model preliminary trial feedback:
Strengths:
● Improved ability to understand prompts, with reduced cases of missing objects and errors;
● The image output relies more on prompts, and the yield of good images has increased significantly.
Weaknesses:
● When the prompt exceeds 400 or 450 tokens, the generated image becomes blurry;
● Multi-character scenes look stiff, character movements tend to be similar, compositions and motions are formulaic with a noticeable AI look. The prompt handling has indeed been enhanced. It would be even better if the aesthetics could match those of anima aesthetics v11. Thanks for the creators' work!
Thanks for the trial notes — this matches what the Q35 stack actually does.
The stronger prompt follow and fewer missing objects come from the native Qwen 0.6B + Qwen 3.5 4B pair. That is the point of this branch. The blur after ~400–450 tokens is a limit of this text-encoder / CLIP path, not a hole in the bake. Past that length the tail of the prompt smears. Shorter prompts hold the frame better than longer ones.
Multi-character stiffness is real. Aesthetic v1.1 was used in earlier versions (V2/V3). On Qwen 3.5 a school shift toward Aesthetic v1.1 breaks the image, so V4 only keeps donors/schools that do not carry Aesthetic understanding. Official Aesthetic v1.1 is a different product; I am not treating it as a target to match on this page. I have not solved that compatibility yet.
Detailer stacks will not fix the token blur or the stiff multi-character look on this line. Clean plate + a style LoRA is safer than extra detailers.
Appreciate the write-up.
@bestluner180 Yes, your model ensures prompt adherence and stable composition, which are the most essential foundational capabilities. Other effects can be gradually improved on this basis or enhanced via LoRA, so it is not a major issue. I tried out 3.8b anima for the first time today and now have an intuitive understanding of it. This is indeed a groundbreaking achievement. Thank you again for your work!