feat(anima): Make Anima respect the precision property in Invoke.yaml

[kappacommit]

Summary

Makes it so that Anima respects the precision property in invoke.yaml

Why?

Anima runs in bf16 precision. This is slow on 2000 series and earlier Nvidia GPUs, and on mac. Allowing the user to configure this, enables that hardware to run the model at adequate performance on those systems.

Why only Anima?

Today - Qwen and Z-Image do not respect the precision property. Which is fine, because neither Z-Image nor Qwen support fp16 at all (it produces broken output.) If we make Z-Image and Qwen respect precision, they will break if set to float16 and if set to float32 they will double their VRAM cost and realistically not be usable on consumer hardware. (Qwen will take >40gb of VRAM)

Anima, SD, and Flux 1 are the only models which safely and feasibly support all precisions. So it is natural to add Anima to the list.

I have documented an issue here for a more robust fix for all models by implementing a similar fix to how Forge and Comfy both handle this automatically. However, testing this is difficult. It requires testing on many different devices (2000 series nvidia gpu, mac): #9153

QA Instructions

1. Generate an image with anima, you will use this as a reference for the other steps
2. In invoke.yaml add precision: float32
3. Generate the same image with anima (same seed / settings). You will notice 2 things: it should take up 2x the VRAM and likely run a bit slower. You will also notice very slight changes to the output image.
4. In invoke.yaml add precision: float16
5. Generate the same image with anima (same seed / settings). You will again notice very slight changes to the image. You likely won't notice much of a VRAM difference this time.

Merge Plan

Standard merge. Backend-only change isolated to Anima paths.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added (4 new unit tests in tests/backend/util/test_devices.py)
• No redux slice migrations needed
• No docs needed (internal helper)

🤖 Generated with Claude Code

[lstein]

I'm seeing more dramatic differences than I expected, and want to double-check that these are within the acceptable range:

No precision in the config file:

precision: float32:

precision: float16:

All were generated with the same prompt, settings and seed.

[kappacommit]

I'm seeing more dramatic differences than I expected, and want to double-check that these are within the acceptable range:
All were generated with the same prompt, settings and seed.

I did another pass to confirm things and it seems this is just an expected quirk of the model. I do also get noticeable differences when switching between them.

There isn't really anything special happening in the code, we just pass the config file value directly to torch and that's it. This is just a configuration option for hardware that can barely run the model otherwise, so they'll likely be happy to make this trade-off.

[lstein]

Change in appearance of model confirmed not-a-bug by the author.
Approved