DEV Community

AI-KSK
AI-KSK

Posted on

Anima 3.8B Enhanced in ComfyUI: 52-Block DiT, Qwen3.5 4B Prompt Understanding and Eight Workflows

#comfyui #ai #tutorial #imagegeneration

Anima 3.8B Enhanced in ComfyUI: 52-Block DiT, Qwen3.5 4B Prompt Understanding and Eight Workflows

Anima has grown from 2.9B to 3.8B, and this release pairs the bigger model with a Qwen3.5 4B text encoder. The notes below come from the episode subtitles and the published mirror page; no workflow JSON was shipped or parsed, so no node-level parameters are claimed beyond what the episode states.

What changed in the model

  • The network was extended into a 52-block DiT and the model size reached 3.8B.
  • The text encoder moved from the old 0.6B encoder to Qwen3.5 4B, which improves prompt following, multi-character binding, character interaction and spatial instructions.
  • Version 1.1 packages the adapter into the model; earlier versions needed a separate adapter.
  • The graph is a 52-block DiT plus a Semantic Connector v2 flow, and the workflow is simpler than before.
  • Community plugins keep compatibility with older LoRAs and ControlNet, and Turbo LoRAs still work for faster generation.

The eight workflows

The mirror bundles eight graphs: standard text-to-image, Turbo acceleration, LoRA, control, repaint, image-to-image and high-resolution upscale routes. The episode walks them in order, starting from the standard generation graph that the others build on.

For the standard non-accelerated graph the episode gives concrete guidance:

  • Set both positive and negative prompts.
  • Because it is not an accelerated graph, CFG can be higher: the official recommendation is 4–7.
  • Step range: 28–50.
  • Sampler/scheduler: Euler + Beta also runs well.
  • Set the resolution, then run.

Deploying the mirror

  1. Open the mirror page: https://www.compshare.cn/images/Khe3CJ9tRb5N?referral_code=A7fTQUYY9lCDFJwA6MMvoq
  2. Deploy an instance — the episode recommends renting a 24 GB or larger GPU.
  3. Wait for the image to come up, then open ComfyUI and pick a workflow from the left panel.

Links

Limits

  • CFG and step ranges above are the episode's recommendations for the standard graph, not a universal optimum.
  • The mirror is a deployment convenience; quality depends on prompt, seed and sampler settings.
  • Third-party LoRAs and plugins keep their own licenses.

Top comments (0)