ComfyUI-Anima-DAVE Introduction
ComfyUI-Anima-DAVE adds DAVE (DC Attenuation for diVersity Enhancement) to a ComfyUI image-generation workflow. It is a single MODEL patch node for Anima / Cosmos-predict2 and MOSS-Image. DAVE is training-free: it changes how the loaded model behaves during generation without requiring you to train a new model.
If you generate several images from the same prompt using different seeds, you may still get nearly identical compositions. DAVE is designed to open up more variation in broad features such as framing, poses, and layout. It works during the early part of generation, when those choices are taking shape.
DAVE is not a guarantee that every seed will produce a different composition. Stronger settings can also make details less reliable, so the best approach is to compare a grid of seeds and adjust the controls gradually.
How ComfyUI-Anima-DAVE Works
Think of each stage inside the image model as carrying two kinds of information: a broad, shared plan for the image and the differences that make one seed distinct from another. DAVE gently reduces part of that shared plan in selected stages. This gives the seed-specific information more room to influence the result.
More precisely, DAVE reduces the DC component—the average value across spatial positions—of selected intermediate model outputs. It leaves the remaining variation, called the AC residual, in place. You do not need to edit these values yourself: the node handles the adjustment while the model generates an image.
DAVE makes this change only in selected model blocks, and normally only during a short early window of denoising. That timing matters. Early changes can influence composition; continuing the edit for too long can damage details such as text and hands.
The node patches the model rather than changing the sampler or an already-generated latent image. Its temporary edits are applied during each model call and removed afterward.
ComfyUI-Anima-DAVE Features
One-node workflow control
Place Anima / MOSS DAVE (after compile) between the model-side nodes and your KSampler:
UNETLoader → Anima / MOSS DAVE (after compile) → KSampler → …
If your Anima workflow uses Anima Block Compile, put DAVE after it:
UNETLoader → Anima Block Compile → Anima / MOSS DAVE (after compile) → KSampler → …
This order is important. If DAVE is placed before Block Compile, the compiled blocks replace the ones DAVE would edit, so DAVE may have no effect.
Controls
| Input | Default | What it controls |
|---|---|---|
mask | auto | Selects which model blocks DAVE edits. auto uses the included mask for the loaded model: blocks 8–18 for Anima or 19–30 for MOSS-Image. |
strength | 0.30 | Sets how much of the selected blocks’ DC component is removed. 0 disables the edit. More removal does not always mean more layout diversity. |
tau | 0.06 | Sets the fraction of early denoising steps during which DAVE is active. At 50 steps, 0.06 is approximately the first 3 steps. 0 means every step, not “off.” |
The included masks focus on blocks associated with broad structure rather than later detail. A custom continuous mask can weight blocks differently; with such a mask, lowering strength can also concentrate the effect on the highest-weighted blocks.
Choosing settings for your artwork
Start with tau=0.06 and strength=0.30. Generate a grid with the same prompt and different seeds, then compare composition as well as detail quality.
- For more varied layouts: Try
tau=0.10and test low strengths first, such as0.05,0.10, and0.20. The layout-diversity effect can peak at these lower values. - For maximum DC removal:
tau=0.10andstrengtharound0.80removes more of the shared signal. This is not necessarily the setting that gives the most distinct layouts. - If text, hands, or objects deteriorate: Shorten
taufirst. The author advises keeping it at or below0.10; at0.15and above, legibility can fall sharply.
A wider time window is generally more damaging than a higher dose within a short window. Raising strength can increase the total change while causing the model to compensate in ways that make layouts converge again.
Compatibility notes
DAVE applies its edit to both the conditioned and unconditioned model passes used by classifier-free guidance (CFG).
The tau control relies on the sampler making its step schedule available to the model. If a custom sampler does not provide that schedule, DAVE falls back to running on every step, as it does when tau=0. If a short-window setting unexpectedly harms detail, this sampler behavior is worth checking.
ComfyUI-Anima-DAVE Models
DAVE is a patch for a model you load; it is not a separate image-generation checkpoint. The node supports these model paths:
| Model path | What DAVE edits | What you may notice |
|---|---|---|
Anima / Cosmos-predict2 DiT | The included Anima mask selects blocks 8–18, an early-to-middle structural range. | More variation in composition across seeds using the same prompt. |
| MOSS-Image | The included MOSS mask selects blocks 19–30 of 42. DAVE edits image-token information while leaving text tokens unchanged. | More varied framing, poses, or panel layouts. Text can remain legible, though hands and objects may be slightly less consistent, and an occasional seed may shift a broad attribute such as sky color. |
For MOSS-Image, the node also excludes 18 unusually large-activation channels identified in its included mask. This prevents those channels from dominating the average DAVE adjusts. These model-specific differences are handled automatically when mask=auto.
For a MOSS-Image workflow using the nodes shown in the author’s example, the model-side order is:
MOSS-Image HF Loader → ModelSamplingSD3 → SkipLayerGuidanceDiTSimple (22) → DAVE → KSampler
The defaults—mask=auto, strength=0.30, and tau=0.06—are also the author’s suggested starting point for MOSS-Image.
Troubleshooting ComfyUI-Anima-DAVE
Different seeds still produce similar layouts
- Check that DAVE is connected on the model path before KSampler.
- If you use Anima Block Compile, check that DAVE comes after it.
- Keep the prompt fixed and compare multiple seeds, so you can judge layout changes rather than prompt changes.
- Try
tau=0.10and sweep lowstrengthvalues (0.05–0.20). Do not assume the highest strength will give the most layout diversity.
DAVE appears to have no effect with Anima Block Compile
Check the connection order:
UNETLoader → Anima Block Compile → DAVE → KSampler
Putting DAVE before Block Compile can cause it to make no edit, even though the workflow runs.
Text or hands become distorted
- Reduce
tauto0.10or below; return to the0.06default if needed. - Lower
strengthand compare another seed grid. - If you use a custom sampler, remember that one without an available step schedule makes DAVE run on every step.
The image changes, but the variation is not useful
Compare composition separately from overall image quality. A high strength can remove more DC without producing better layout diversity. Try the lower-strength range (0.05–0.20) before increasing it.
Frequently asked questions
Does strength=0 turn DAVE off?
Yes. It disables the DC-removal edit.
Does tau=0 turn DAVE off?
No. It applies DAVE at every denoising step. Use strength=0 to disable the edit.
Will DAVE change the text part of a MOSS-Image block?
No. For MOSS-Image, the edit is calculated from image tokens; text tokens pass through unchanged. Generated text can still be affected indirectly by broader changes to the image.
Is a higher strength always better?
No. Low settings can produce the strongest layout diversity. Higher settings remove more of the shared signal but can reduce legibility or consistency.
Learn More about ComfyUI-Anima-DAVE
- DAVE paper repository — background on the method introduced in Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation.
- MOSS-Image repository — information about the MOSS-Image model and its ComfyUI workflow.
ComfyUI-Anima-DAVE is MIT-licensed. The DAVE method is credited to its authors.
