DEV Community

AI-KSK
AI-KSK

Posted on

Anima Update Burst: Image Editing LoRAs, Four-Step DMDX and Artist Mixing

#comfyui #ai #tutorial

Five things land in one episode: Anima can now be edited with LoRAs, a DMDX student model cuts sampling to four steps, artist tags can be mixed in cross-attention instead of being stuffed into one prompt, a developer's Flow sampler shows up, and the community's second-order latent upscale gets a workflow. The author also corrects his own earlier mistake about the inpainting model.

Evidence note: three workflow JSONs shipped, but only two can be tied to the flows the video shows - see the note in the parameter section about the file named Anima Edit 图像编辑工作流.json. Edit-test verdicts are the author's, as reported in the video.

When to use it

  • You want Anima to do prompt-driven edits, not just generation.
  • You need speed: 4 steps at CFG 1, near-realtime at 512x512.
  • You want controlled style mixing between artists without producing a four-way blur.

What is different in this workflow

  • Three editing routes compared: Anima LLLite Inpainting (an inpainting model the author had been using wrong - it is meant to be wired like an IC-LoRA, not as a plain local repaint), AnimaEdit, and a third LoRA. For large-area edits the author says only LLLite works.
  • DMDX is a student model, not a LoRA drop-in: 4 steps, CFG 1, and the author reports it keeps style surprisingly well.
  • Artist mixing is a separate node: artist tags are encoded separately and mixed at the cross-attention layer, with a layer window (28 layers, max 27) to control where the mixing applies.
  • The Flow sampler is a purpose-built path around Anima's flow logic, not a generic sampler swap.
  • Second-order upscale uses the latent, which is the author's standing conclusion.

Step by step

  1. DMDX: load the Anima UNET, dmdx_student_final_comfy LoRA at 1.0, 4 steps, CFG 1, er_sde + simple. At 512x512 the author calls it near-realtime rerolling.
  2. Editing: pick a route deliberately. Large-area changes -> LLLite inpainting; prompt-style edits -> AnimaEdit. Test set in the video: jacket to faux fur (LLLite succeeds), ice cream to pink with sprinkles (Edit succeeds), background to a coffee shop (LLLite), red evening gown (LLLite), pink hair (nothing fully succeeds), pearl necklace and earrings (Edit closest).
  3. Artist mixing: @artist on top, prompt below, negative prompt after; keep the other options at default. The node was not yet supported on RunningHub, so the author ran it locally.
  4. Layer window: 28 layers, so the max is 27/-1; set a start and end if you want the mixing to stop early (his example: 6).
  5. Flow sampler: default flow path; the author reports at minimum no degradation.
  6. Second-order upscale: generate larger, but keep the element/style retention of the 1 MP stage.

Parameter basis

  • DMDX (Anima DMDX 四步极速版.json, 11 nodes): UNETLoader -> anima-base-202605-v1.0.safetensors; CLIPLoader -> qwen_3_06b_base.safetensors (type stable_diffusion); VAELoader -> qwen_image_vae.safetensors; LoraLoaderModelOnly -> dmdx_student_final_comfy.safetensors at 1.0; ModelSamplingAuraFlow; KSampler -> 4 steps, CFG 1, er_sde, simple.
  • HiresFix (Anima HiresFix 1.5倍精修.json, 12 nodes): anima_baseV10.safetensors; two KSampler at 34 steps CFG 4.5 er_sde simple (denoise 1.0) then 18 steps CFG 4.5 denoise 0.35; LatentUpscale.
  • The third file (Anima Edit 图像编辑工作流.json, 112 nodes) contains LTX-2.3 audio/video nodes (LTXVImgToVideoConditionOnly, LTXVSeparateAVLatent, SaveVideo) and an LTX checkpoint, not Anima editing nodes - its contents do not match its name, so this article does not quote parameters from it.

Limits

  • Material defect: the file named Anima Edit 图像编辑工作流.json is an LTX2.3 audio-video graph, so the editing flow the video demonstrates cannot be audited here.
  • Editing is uneven: the author's own six-case test has a complete failure (pink hair) and a case where clothes got polluted.
  • The LLLite inpainting model is not an edit model; using it as plain local repaint was the author's earlier mistake.
  • Artist mixing was not supported on RunningHub at the time, so it is a local-only path.
  • DMDX is a student model and the author does not dig into how it was distilled.

Episode resources

Top comments (0)