Checkpoint merge on Anima-2.9B (40-block). Same idea as the 3.8 file: anime / illustration first, some kemonomimi. Use it like a normal Anima model.
Text encoder: qwen_3_06b
VAE: official Anima / Qwen-Image VAE
Prompt: anime, masterpiece, best quality, ultra detailed, absurdres, sharp focus, intricate details, cinematic dramatic lighting, strong chiaroscuro,
V5 — digital art. Gloss, hard contour, 2.5D texture.
V6 — Aesthetic line, same family as V2/V3. Flatter anime drawing, rougher paint. Not a clarity upgrade over V5 and not a return to it.
Ksampler
Steps: 30/40
CFG: 4.0–5.5
Sampler: Euler ancestral
Scheduler: normal
detail / upscale
Steps: 10–20
CFG: 2.0–3.5
Sampler: DPM++ 2M
Scheduler: sgm_uniform
These files are 40-block Anima-2.9, same chassis as the original 2.9B. Stock Comfy often loads them as 28 blocks with no hard error — that is the colored static.
You need the loader patch from the original 2.9B author:
https://github.com/gazingstars123/ComfyUI-Anima-2.9B
Original 2.9B page (same depth expansion):
https://civarchive.com/models/2855007/anima-29b
Put the folder in custom_nodes, restart, then check the startup/load log for 40 blocks, not 28.
Also still need: official Anima / qwen_image VAE, CLIP qwen_3_06b_base + qwen_image, 16ch latent (not EmptyLatentImage 4ch), AuraFlow shift ~3–4.5. Delta Mix nodes are not required just to run the baked checkpoint.
Credits
Anima Base / Anima-2.9B — original authors
Method: Anima Delta Mix — 2B into 2.9B / 3.8B
V1 - Anima Base / Anima-2.9B: G1 - Aura Rising
V2 - Anima Base / Anima-2.9B: Stem CC-Dream Island-anima v2.0/ G1 aniMurr
V3 - Anima Base / Anima-2.9B — Stem Chosen-mix_anima / G1 aniMurr
V4/V5/V6/V7 - Anima Base / Anima-2.9B: Custom Stem/G1
V7.5 - Direct evolution of v7. Increased base image sharpness, cleaner generation. Polished variant focused on clean lines and frame clarity.
License follows Anima (non-commercial). Do not sell the weights.
Description
FAQ
Comments (3)
When using the model Anima Delta Mix 2.9B v3.0, the LoRA doesn't seem to take effect, but no errors are reported at runtime. However, when using Anima-2.9B, these same LoRAs work fine. Does this model need a new node to load LoRAs?
Official 2.9 often remaps old 28-block LoRAs through the Anima pack.
v3 is a 40-block bake; stock Load LoRA can attach to nothing and stay silent.
Need:
1. Which loader — Load LoRA or a remap node?
2. Were those LoRAs trained on 28-block Anima or on official 2.9 (40)?
3. Console after Queue: any “lora loaded / missing keys”?
If they are 28-block LoRAs, use the Delta Mix 2.9 / Anima-40 remap
(same pack as the mix nodes), not stock Load LoRA.
Do not use the 3.8 / 52 LoRA node on this file.
If the log shows 0 keys applied, it’s the map, not a dead bake.
@bestluner180 Fixed, the remap node solved it. Thanks for the explanation!