WAI-ANIMA is a merge built on Anima Preview 3, and this episode is less about the model than about where to do the upscale. The author runs the comparison properly — pixel space versus latent space — and the latent path keeps more detail. The second finding is a ceiling: past roughly four million pixels the style starts to collapse and artefacts bloom.

Evidence note: three workflow JSON files ship with this episode and were parsed directly; node counts, UNET/CLIP/VAE filenames, LoRA strength, latent sizes and the sampler family below come from those files.

When this is useful

  • Anime and illustration generation where a merged Preview 3 model suits the style you want.
  • Getting from a soft 1M-pixel draft to a clean final without the style flipping.
  • Tiled upscale of an existing image up to about four million pixels.
  • Image-to-image with control over tone and composition at a known denoise.

How this differs from the previous build

Three findings matter more than the model itself.

  • Latent-space upscale wins. The graph that does not decode before upscaling produces visibly more detail than the pixel-space path — the author had to put them side by side to see it.
  • One million pixels for the first pass. Above that, both aesthetics and stylisation degrade, which is why the earlier 1.3M test grew excess texture on hard surfaces like pillars.
  • A special sampler with an eta value. At eta 0.35 the results stay stable; with an ordinary sampler the same prompt produced broken limbs.

Running it

  1. Text-to-image: write the positive prompt, and the negative too — Anima is not distilled, so the negative actually works. Use the model author's recommended negative.
  2. Keep the first pass at about one million pixels and set the ratio.
  3. Run a hires-fix style pass: upscale in latent space by 1.5x, then sample again.
  4. Image-to-image: load the image, constrain the size, set the prompt and denoise — 0.73 in the demo.
  5. For style control, prefix the prompt with @artist tags. Changing artist changes the style dramatically; changing the aspect ratio changes it again.
  6. Tiled upscale: load the image, set the size, tag it with the lightweight WD14 tagger for a prompt, and add the artist prefix so the style survives.
  7. Use the special sampler with eta 0.35 for the upscale sampling pass.

Parameters, and where the numbers come from

Item Value in the shipped graphs
UNET (all three graphs) waiANIMA_v10（anima3）.safetensors, weight dtype default
Text encoder qwen_3_06b_base.safetensors (qwen_image; one graph declares stable_diffusion)
VAE qwen_image_vae.safetensors
T2I latent 512x512, batch 1
Character LoRA (tiled graph) Gundam_RX78_Flux.safetensors at model 0.85 / clip 0.85
Sampler family ClownsharKSampler_Beta in all three graphs
T2I first pass eta 0.35, linear/euler, beta57, 30 steps, denoise 1, CFG 3
T2I hires pass eta 0.35, linear/euler, beta57, 30 steps, denoise 0.55, CFG 3
I2I pass eta 0.35, linear/euler, beta57, 30 steps, denoise 0.5, CFG 3
Tiled upscale pass eta 0.5, linear/euler, beta57, 20 steps, denoise 0.3, CFG 2
Upscale approach LatentUpscaleBy 1.5x in the T2I graph; ImageUpscaleWithModel plus tile nodes in the tiled graph

The eta values are worth reading closely: the graphs ship 0.35 on the generation passes and 0.5 on the tiled-upscale pass, which matches the narration. The one place the video and the file disagree is the image-to-image denoise — the narration quotes 0.73 while the shipped graph stores 0.5.

The Gundam_RX78_Flux LoRA appears in the tiled-upscale graph as a character demo — it is not part of the base recipe, so treat its 0.85 strength as an example of LoRA loading rather than a required setting.

Honest limits

  • The one-million-pixel and four-million-pixel limits are empirical, from the author's own tests, not model documentation.
  • Tiled upscaling does lose some detail even when it sharpens the image; the author expects a later model version to fix it.
  • Losing the artist prefix on an image-to-image pass drops the style hard — the prefix is doing real work.
  • eta 0.35 is tuned for stability, not for maximum beauty; the author says the result is stable rather than beautiful.
  • Anima Preview 3 can still break anatomy at high resolutions, which is what the special sampler is compensating for.

Episode resources

RunningHub workflows