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: 30-40
CFG: 4.0–5.5
Sampler: Euler ancestral/dpmpp_2m
Scheduler: normal/sgm_uniform
Resolution: same as Anima-3.8B (1024-class and up, as VRAM allows).
Anima 52 LoRA Patch only for v1
https://github.com/bestluner-create/ComfyUI-Anima52-LoRAPatch
Anima 52 LoRA Patch for v2+
https://github.com/bestluner-create/ComfyUI-Anima-Delta-Q35-LoRA.git
Start around 0.65. Do not stack stock Load LoRA on the same old LoRA.
v2.0
Current version on this page.
This mix uses: Anima-3.8B Anima Aesthetic 2B Aura Rising aniMurr v02
Same text encoder and VAE as v1.0. File: animaDeltaMix38B_v20.safetensors
v3.0
New bake on the same 52-block 0.6B chassis
Both Gen-1 and Gen-2 use copy-body (adaLN stays on the chassis)
Denser materials and small surface detail vs v2
Same 2B donors — last styled bake on this pair
v4.0
Custom models are used for both generations, including the stem. This specific model focuses on delivering a high-quality output with maximum detail and fidelity. I hope you enjoy it.
Credits
Anima Base / Anima-3.8B — original authors
Method: Anima Delta Mix — 2B into 2.9B / 3.8B
License follows Anima (non-commercial). Do not sell the weights.
Description
v3.0
New bake on the same 52-block 0.6B chassis
Both Gen-1 and Gen-2 use copy-body (adaLN stays on the chassis)
Denser materials and small surface detail vs v2
Same 2B donors — last styled bake on this pair
Sampler unchanged: 40 steps, CFG 6, scheduler normal. Same-model 2nd pass at 1.25 if you want a tighter frame.
FAQ
Comments (46)
I couldnt make this work, i downloaded the correct clip model, i already had the qwen vae, i am trying to run the model with simplest possible workflow and i only get dark static.
Same. Pure static, no matter which workflow I use, sampler, or scheduler.
dark static usually means the graph is running this as SDXL / EPS, not as Anima.
This file is Anima 3.8B FLOW + qwen_3_06b only. Pony / Illustrious / Qwen 3.5 workflows will just give noise.
Easiest path: open any preview on this page, use Create or copy the nodes under the image. The PNGs have the workflow in metadata, so you can drag the image into Comfy and rebuild from that.
If you build it by hand:
UNET - this bake
CLIP - qwen_3_06b_base (not the 4B)
VAE - official Anima / Qwen-Image
Euler / Euler a, normal, 30–40 steps, CFG 4–6
Shift around 3–4.5 if your sampler has it
Shift 0 or the wrong CLIP = black static. v3 is 0.6B, no Connector, no Qwen 3.5.
@bestluner180 Unfortunately, I tried your workflow as well, with every single part you had. Still static. Not saying it doesn't work, just not for me. And it is a shame, because I LOVE that artstyle.
@Ellen_Weiss Got it - if that graph still gives static, I need the log, not another sampler guess.
Please Queue once and paste the console text from “got prompt” down to the error / first sampling line. Especially:
- model_type (FLOW or EPS)
- which CLIP file loaded
- which VAE
- any red error
Also say whether the model is in diffusion_models / unet or in checkpoints.
Without that I’m guessing. The bake is fine; something in the load path is still not Anima FLOW.
@Ellen_Weiss If the console is messy, just export the workflow.
In Comfy: Save (API format) and drop the json here,
or drag your working graph PNG.
That shows the loader, CLIP, VAE and Shift in one file.
Faster than guessing from static.
@bestluner180 I have the API, but would you be so kind as to instruct me on how to upload it here?
@Ellen_Weiss CivitAI comments don’t like raw json. Either:
1. Rename the file to .txt and attach it if the comment box allows a file
2. Or open the json, copy everything, paste into https://pastebin.com and drop the link here
3. Easiest: in Comfy save the output PNG and upload that image to this comment — the workflow sits inside the picture
Pastebin or PNG is enough. Don’t paste the whole json into the comment body.
https://pastebin.com/zUdqMLbq Would this work, @bestluner180 ?
@Ellen_Weiss Found it. The bake and CLIP are fine.
Your graph uses EmptyLatentImage — that is 4-channel SDXL latent.
Anima needs a 16-channel latent (EmptySD3LatentImage, or the empty latent from an official Anima / Qwen-Image workflow).
Also add ModelSamplingAuraFlow (or SD3) with Shift 3–4.5 on the UNET before the sampler.
For a clean test: bypass Lora Manager, UNET → Shift → KSampler → VAE Decode.
Same 0.6B CLIP and Qwen-Image VAE you already have.
4-channel latent = black static even with the right model. That’s why previews on this page work and yours doesn’t.
@bestluner180 Don't think that was it, or maybe I did it wrong. This is the new workflow, still static: https://pastebin.com/chraPGzA
@Ellen_Weiss Closer. EmptySD3 + Shift 4 are correct.
Two things still wrong in this json:
1. Lora Manager is still between the model and the sampler.
Bypass it for the test. Wire:
UNET (or Checkpoint loader) → ModelSamplingAuraFlow → KSampler → VAE Decode
No Lora Manager.
2. You’re still on UNETLoader. Put the same file into CheckpointLoaderSimple
like the previews on this page.
3. The second KSampler has no model input at all. Ignore that branch.
If the first Save Image is still static after bypassing Lora Manager + Checkpoint loader, then paste the console from “got prompt” (model_type line).
@bestluner180 Nope. I tried, no luck. Then I made a new, simple, direct workflow. Also no luck: https://pastebin.com/FnEeff2b
@Ellen_Weiss This json is simple, but the loader is wrong.
animaDeltaMix38BQwen0_deltamix38BV30.safetensors is a UNET / diffusion_models file, not a full checkpoint.
Checkpoint Loader (pysssss) treats it as SD and you get EPS + static.
Use this only:
Load Diffusion Model (UNETLoader)
file in models/diffusion_models
→ ModelSamplingAuraFlow shift 3–4.5
→ KSampler (normal, not Advanced)
CLIPLoader qwen_3_06b_base type qwen_image
VAELoader qwen_image_vae
EmptySD3LatentImage
No pysssss Checkpoint Loader. No ModelSamplingSD3.
If UNETLoader does not list the file, move it out of checkpoints/ into diffusion_models/ or unet/ and refresh.
Comfy has to be a build that already runs official Anima. If official 2.9/3.8 0.6B also static on your PC, the issue is Comfy, not this bake.
Ok apparently the issue was the block size reading from model loader nodes, installing anima 3-8b custom nodepack fixed the issue.
@bestluner180 Yeah, I changed it. I tried other models I have for Anima, including the official one. They all work, but not yours. Sorry, I'm throwing in the towel. AMAZING art, though.
@Ellen_Weiss My miss — I should have asked for the Comfy console first, not the json.
If the log loads 28 blocks on this 52-block file, the loader pack is missing.
Stock UNETLoader does not know 40/52. Rootsyl fixed the same static by installing
comfyui-anima-3-8B (the official 3.8 pack). The 2.9 pack may be enough for 40,
not for 52.
After install: restart Comfy, then check the console on Queue.
You want 52 blocks + FLOW, not 28.
Workflow can look perfect and still noise if the depth patch is not there.
@Ellen_Weiss No worries. Official 2B / 2.9 can run without the 52-block pack. This file is 52.
Last check if you still want it: on Queue, does V3 log 52 blocks or 28?
File size should be ~6.99 GB. If the console says 28, the 3.8 pack is still missing.
If official 3.8 (52) works and only this bake static — say so. Then it’s the file, not your graph.
Glad you like the previews. Thanks for posting the json.
@bestluner180 I have been getting a headache. It does say 28, but the node does not want to install no matter what. I downloaded it much like other nodes, updated comfy (desktop version), but no luck.
@Ellen_Weiss 28 blocks = the 52-file is still read as 28. That’s why it’s static.
Comfy Desktop is picky with this pack. Try in order:
1. Comfy Manager → search “anima 3.8” / comfyui-anima-3-8B → Install → restart Desktop
2. If Manager says installed but still 28: the folder is the wrong custom_nodes.
Desktop uses its own path (not the portable ComfyUI/custom_nodes).
Clone the pack into that Desktop custom_nodes folder, restart.
3. If it still won’t load — use Comfy portable, not Desktop. Official Anima 3.8
people mostly run portable.
Until the console says 52, no sampler tweak will help. Not your fault and not a dead bake.
update your comfyui to newest if you get dark static
@fre19861 It is updated. Like @bestluner180 said, it was just comfyUI desktop being somewhat incompatible.
@Ellen_Weiss Glad it clicked. Hope it runs as intended from here enjoy the checkpoint.
Qwen 3.5版より破綻が少ない印象
でも3.5の方が出力イメージは素直で悩む
翻訳機で書いています。意味がずれていたらすみません。
いくつか違うプロンプトで両方試してみてください。
場面によって出方が変わります。
どちらか一つに決めなくても大丈夫です。両方使えばいいと思います。
@bestluner180 出力される絵柄が好みで使わせてもらっています、ありがとう
Hi, thank you for creating this model! I’m using ANIMA as my base, and I’ve been seeing a lot of architecture expansions recently, like 2.9B, 3.8B, etc. So I have a couple of questions:
1. If I want to use this, do I need to update ComfyUI and the custom nodes to match the new architecture, and patch the LoRA blocks every time?
2. I’ve been seeing these expansions happening more and more often. I saw the 2.9B version last month, and now there’s already a 3.8B version, and previous LoRAs broke with these changes. If the trend continues and the architecture keeps expanding, potentially to 12B or even larger, I’m not sure that’s a good sign, especially for people who rely heavily on LoRAs.
I’m curious what you think about this direction. Do you expect the architecture to keep expanding like this in the future?
Thanks for writing.
1. You only update for the file you load.
2.9 = 40 blocks, 3.8 = 52. Comfy needs a build / Anima pack that actually reads that depth
(console should say 40 or 52, not 28). That’s a one-time pack, not a patch every bake.
Old 28-block LoRAs do not land on 40/52 with stock Load LoRA. Official 2.9 sometimes remaps
them; this mix often does not. Use the remap node for that depth (40 ≠ 52).
LoRAs trained on the same 40/52 usually load normally.
2. I don’t make the chassis. 2.9 and 3.8 are official insert floors, already in the box.
Delta Mix only writes 2B donors into those slots. I can’t promise 12B or that LoRAs will
keep working across every future depth.
If you live on LoRAs, stay on one depth (2B or 2.9 or 3.8) and train / remap for that
depth. Chasing every new floor will keep breaking 28-LoRAs. That’s the cost of inserts,
not a sign this mix will grow by itself.
The model isn't working; I'm getting a black image or a mess of multicolored noise. I used the parameters specified in the model description, but nothing works.Comfy console first, not the json.
If the log loads 28 blocks on this 52-block file, the loader pack is missing.
Stock UNETLoader does not know 40/52. Rootsyl fixed the same static by installing
comfyui-anima-3-8B (the official 3.8 pack). The 2.9 pack may be enough for 40,
not for 52.
After install: restart Comfy, then check the console on Queue.
You want 52 blocks + FLOW, not 28.
Workflow can look perfect and still noise if the depth patch is not there.
Hi, I partially resolved the issue and updated ComfyUI to version 0.35.0, and your 2.9-billion parameter model is now working. However, I prefer using SwarmUI because I find it more convenient and intuitive. I’m wondering if the 3.8-billion parameter Anima model—specifically the version for Qwen3 0.6B—will also work. Does the 3.8B version require additional nodes, or does ComfyUI have native support for it, just like it does for the 2.9B version?I installed the nodes in `custom_nodes` for ComfyUI and `DLNodes` for SwarmUI just to be safe, and I set up the following: ComfyUI-Anima-2.9B-loraPatch, ComfyUI-Anima-3.8B-LoRA-Bridge, ComfyUI-Anima-3.8B-loraPatch, ComfyUI-anima-3-8B, ComfyUI-Anima-Delta-Q35, ComfyUI-Anima-LoRA-ControlNet-Patch, ComfyUI-Anima-Remap-1.1.0, comfyui-custom-scripts, and comfyui-easy-use. I did this after updating ComfyUI itself; once I updated ComfyUI within SwarmUI, the Anima 2.9B model started working—I even tested it with older LoRAs meant for the Anima 2B model, and everything seems fine. However, I’m unsure about downloading the 3.8B version. I’ll just reiterate that I’m not an expert on ComfyUI; I mostly use SwarmUI because it makes sense to me and doesn't involve a confusing mess of nodes.I installed the nodes I shared earlier, just in case—I’m not sure if SwarmUI will pick them up or not. I installed ComfyUI-GGUF similarly before, and it worked without the nodes; I was simply able to use .gguf models right away after installing the GGUF node.I tried to set up the workflow, but when I looked at it, I barely understood anything; plus, half the nodes were red and labeled "unknown." I loaded `workflow_1.1.json` and `workflow_2_pass.json`, but—as I mentioned before—I don't really know what I'm doing; I just loaded the model and hit "Generate" without realizing I needed to update ComfyUI from 0.28.0 to 0.35.0 or that specific nodes were required. Also, it seems those workflows require nodes designed for RTX-series cards, whereas I have a GTX card.If ComfyUI and SwarmUI can run without nodes, I’ll use Anima Delta Mix 3.8B Qwen 0.6B; if not, I’ll stick to the 2.9B version. And if the Anima Delta Mix furry edition 2.9B Qwen 0.6B becomes available, I’ll download it when I have time.I don't want to get bogged down in the complexities of nodes, as I might break ComfyUI or SwarmUI; it's easier for me to just download the model, configure the generation settings, and start generating.@Roozymeko Hi. You don’t need that node pile for this file.
Anima Delta Mix 3.8B Qwen 0.6B is the same family as 2.9: diffusion model + Qwen 0.6B + VAE. Current ComfyUI (you already have 0.35) can load 52-block Anima without extra architecture nodes.
Do not install for this checkpoint:
ComfyUI-Anima-Delta-Q35
ComfyUI-Anima-3.8B (Q35 / Connector / 4B path)
Remap 1.1
the workflow_1.1.json / dual-encoder graphs
Those are for the Qwen 3.5 page, not this one. Red “unknown” nodes are from that graph. GTX vs RTX is unrelated; those workflows are just the wrong pipeline.
SwarmUI: if 2.9 already generates after the Comfy update, try 3.8 the same way: pick the 3.8 file, Qwen 0.6B encoder, same sampler. No custom node required for a plain generate.
LoRA: old 2B / 2.9 LoRAs on 3.8 are optional. Only add a LoRA patch if a LoRA actually breaks. Don’t stack every patch “just in case”.
If Swarm still refuses 3.8 after that, stay on 2.9. That is the simple path you want. Furry 2.9 is a separate page when it exists — this 3.8 file is not that.
Hi, I tested your 3.8-billion parameter anime model (using the Qwen3 0.6B encoder) and it works—I used these settings: Sampler: Euler Ancestral, Scheduler: beta_1_1 (new), 20 steps, and CFG scale 5. I also tested it with LoRAs—specifically LoRA sliders—and haven't noticed any bugs so far. It would be a good idea to mention in the descriptions for your 2.9 and 3.8 models that the latest version of ComfyUI is required. The developers might have added native support in ComfyUI version 0.35.0; since I use SwarmUI (which runs on top of ComfyUI), it’s likely the ComfyUI developers fixed the incompatibility issues affecting the 0.6B encoder for the 2.9 and 3.8 models. That’s just my speculation, of course, but it is a fact that the 2.9 and 3.8 models started working for me only after I updated ComfyUI to version 0.35.0.It would be great if you added a note to the descriptions of your 2.9 and 3.8 models (based on the Qwen3 0.6B encoder) stating that ComfyUI version 0.35.0 is strictly required. You should explicitly mention—perhaps by citing the older version 0.28.0 as an example—that using an outdated version will likely cause issues. Since SwarmUI depends on ComfyUI, this implies that models 2.9 and 3.8 won't work there either unless ComfyUI is updated to the current 0.35.0 version.It’s possible that both SwarmUI and ComfyUI are using the nodes located in the folders I sent you earlier; there’s a chance SwarmUI picked them up, which might be why your 3.8 model is working—I didn't have those nodes in my ComfyUI `custom_nodes` or SwarmUI `DLNodes` folders before. I’m not really an expert on this, though, so I don't know exactly how SwarmUI detects or loads the nodes needed to run the 3.8 model.@Roozymeko Good that 3.8 + Qwen 0.6B runs after the Comfy update.
This file is a normal Anima 52 / 0.6B load. You don’t need the Q35 / Remap / extra LoRA packs just to generate. If Swarm works, it’s because Comfy behind it is current.
I’ll add a short note on the 2.9 and 3.8 Qwen 0.6B pages: use a current ComfyUI (0.35 or newer). Old 0.28-class builds are a common fail.
My sampler notes on the card stay Euler a + normal for the sketch and DPM++ 2M + sgm_uniform for refine. Your beta_1_1 / 20 steps test is fine if the image holds.
I tested the generation process myself without the custom nodes and was pleasantly surprised; everything runs smoothly on SwarmUI (which operates on top of ComfyUI 0.35.0). You were right to suggest removing those nodes—some are unnecessary, while others might be outdated and could potentially break or degrade the generation quality. To sum up, the update to ComfyUI 0.35.0 completely resolved the issues with the Anima 2.9 and 3.8 models, which require the Qwen3 0.6B encoder. However, I’m not sure if models requiring the Qwen3.5 3.8B–5B encoder will work without those nodes, as I haven't downloaded or tested them personally. As for the LoRAs, I’ll download the patch; it certainly won't hurt, and it's better to be safe.Thanks a lot for helping me figure out which nodes are necessary and which ones are better to remove for stability. I was also happy to help test it on my PC—here are the specs if you're interested: Intel(R) Core(TM) i5-7500 CPU @ 3.40GHz (4 CPUs), ~3.4GHz, NVIDIA GeForce GTX 1050 Ti 4GB VRAM, 16GB RAM. On average, generating a single image takes 10–15 minutes, which works well for me; I had already done quite a bit of optimization in ComfyUI for my machine. (Please excuse any translation errors; I used Google Translate.)I just want to give an update on this particular model.
While the newer models provided a lot of things that I can appreciate, however, it did stray away far from the vibrant colors that V1.0 had and I was really looking forward on how it would be better improved.
Sadly, the colors have dulled quite a bit from V1.0
Yes, this is related to aesthetics. In the next version, I plan to find a balance between aesthetics and color.