Anima 3.8B for Forge Neo

A Forge Neo extension that runs the Anima 3.8B diffusion model with its Qwen3.5 4B conditioning path, including the fixed-strength timestep-aware Semantic Connector v2 release.

Bundled v2 checkpoints activate automatically. The adapter dropdown and strength controls remain available only as a backup for legacy v1 checkpoints.

Requirements

• Forge Neo neo 2.28 or newer.
• The Anima checkpoint must already work natively in your Forge install (that is where qwen_3_06b_base.safetensors and qwen_image_vae.safetensors come from — this extension does not replace them, it adds a second encoder).

Install

1. Install the extension

git clone https://github.com/GumGum10/forge-anima-3.8B extensions/sd-forge-anima-3-8b

Run that from your sd-webui-forge-neo folder, so you end up with:

sd-webui-forge-neo/extensions/sd-forge-anima-3-8b/

There is nothing to pip-install; the Qwen3.5 tokenizer is bundled in qwen35_tokenizer/, so the extension never touches the Hugging Face cache or the network.

2. Download the models

For v2, place the combined checkpoint in Forge's checkpoint directory:

The older separate v1 checkpoint and adapter remain supported:

These two you already have from native Anima — leave them where they are:

models/text_encoder/qwen_3_06b_base.safetensors
models/VAE/qwen_image_vae.safetensors

All paths are relative to your sd-webui-forge-neo folder. If you also run ComfyUI, hard-link the big files instead of copying them:

mklink /H "models\Stable-diffusion\Anima-3.8B-v2.safetensors" "C:\path\to\ComfyUI\models\diffusion_models\Anima-3.8B-v2.safetensors"

Restart Forge after copying the files.

Use

1. Select the Anima-3.8B-v2 checkpoint.
2. In Forge's VAE / Text Encoder control, select qwen_image_vae and qwen_3_06b_base, exactly as you would for native Anima.
3. Generate normally. Forge detects the bundle metadata and activates v2 even when the extension accordion is collapsed.
4. Optionally expand Anima 3.8B (Qwen3.5 / v2) and enable Use adapter on negative prompt. V2 always uses its trained strength of 1.0.

For a legacy v1 checkpoint, enable the accordion, select its separate adapter, and choose the desired strength.

The positive prompt is encoded by both text encoders and fused by the selected adapter. The negative prompt stays on native Anima — the reference workflow's behaviour — unless you enable the negative switch, in which case it takes the same path at its own strength.

Everything is restored after each generation, so disabling the accordion gives you stock Anima again without a restart.

Notes

• V2 bundles are discovered by safetensors metadata, not by filename.
• Legacy separate adapters use architecture anima_progressive_qwen35_cross_adapter_v1.
• The Qwen3.5 model is found automatically; its filename must contain qwen35_4b, qwen3.5-4b, or qwen3_5_4b.
• The native encoder and Qwen3.5 are run sequentially and released after their outputs are copied to system RAM. On lower-VRAM systems, Forge can offload text-encoder and connector weights and reload them when a prompt changes.
• Semantic Connector v2 remains active during every denoising step because its extraction is timestep-aware. Forge manages it as part of the sampling model and can stream/offload its layers when the full model does not fit.
• Adapter name, strength, and negative strength are written into the generation parameters of every image.

Bundling a newer v2 epoch

The included bundle_v2.py validates that the new adapter was trained against the native adapter inside the selected DiT, then writes a new combined checkpoint. It deliberately refuses to overwrite an existing bundle:

python bundle_v2.py `
  --base "path\to\Anima-base.safetensors" `
  --adapter "path\to\expanded_adapter_v2_epXX.safetensors" `
  --output "models\Stable-diffusion\Anima-3.8B-v2-epXX.safetensors"

Use a new output filename for every release epoch, refresh Forge's checkpoint list, and select the new bundle. No code change is needed while the adapter architecture and native-adapter hash remain compatible.