Pinokio
FeedApps
Download Pinokio
Sign inCreate accountJoin
Sign inCreate accountJoin

Post

Conversation

HiDream O1 Full FP8 Support

@cocktailpeanut
5mo ago

Just added support for FP8-FULL (previously only FP8-Dev was supported!)

Commit: https://github.com/cocktailpeanut/hidream-o1/commit/ce9b77f4b4b4a59ae350016ceb207a303551d593

Main reason

In many cases there is not much difference between full vs. dev. (except that full is slower). But we found a problem where image editing sometimes gave us faded images like this:

when it should look more vivid like this:

And turns out, this is a known issue with the dev model. So for image editing, I recommend using the full model (although it's slower it's def worth it)

VRAM

They seem to use the same amount of VRAM.

Difference

  • FP8-dev
    • uses 28 steps for generation. Faster.
    • But has trouble with image editing---generates faded images.
    • illustrations/anime work better on this model
  • FP8-full
    • uses 50 steps for generation. slower.
    • The image editing issue doesn't exist. no faded images
    • photography/realism images work better
HiDream O1 Image FP8cocktailpeanut/hidream-o1
0

Replies

0
Up to 10 files, 25MB each. Images are optimized; GIFs -> MP4; videos 720p (max 120s).

Post context

App used

HiDream O1 Image FP8cocktailpeanut/hidream-o1

One-click launcher for the original HiDream-O1-Image web UI using lazy-downloaded drbaph Dev or Full FP8 checkpoints through a root FP8 runner. Req...

View app

Published by

@cocktailpeanut5mo ago
Pinokio
PrivacyTerms
FeedApps

Post context

App used

HiDream O1 Image FP8cocktailpeanut/hidream-o1

One-click launcher for the original HiDream-O1-Image web UI using lazy-downloaded drbaph Dev or Full FP8 checkpoints through a root FP8 runner. Req...

View app

Published by

@cocktailpeanut5mo ago

Related posts

HiDream-O1-Image: a single transformer image generation model with reasoning skills@cocktailpeanut · 5mo agoRun HiDream-o1-image with Low VRAM (10GB)@cocktailpeanut · 5mo agoWhy YUE2 // GROOVE keeps YuE2's reference configuration — and what "16 GB" really means@bearbone · 25d ago