← sources

Possible flaw in H3 Image Edit default workflow: 40% faster runtime

source: reddit (reddit:StableDiffusion)published 2026-11-11 07:45ingested 2026-11-12 02:40

retained text

On my 5060ti, H3 image edit went from 263s to 158s based on the following rationale.

The default H3 Image edit workflow has MP adjustment at the Resize Image/Mask node set immediately after the input image which then feeds 3 things:

1. The prompt writer
2. The image conditioning node
3. The width/height of the image conditioning node which determines output MP.

In that setup, a user who specifies 16MP is sending a 16MP image to the prompt writer and the conditioning node, which is unnecessary overhead and potentially producing noise.

I altered the workflow in two ways:

1. Fed the input image directly to the prompt rewriter (I suggest downsizing to 1MP if your input image is large)
2. Fed the input image through a separate chain of two Resize Image/Mask nodes set to 2MP and then multiple of 32, just like the original chain.
3. Leave the original chain supplying output width/height in place, at 16MP for output dimensions.

Edit:

There is a concern regarding the potential loss of not feeding a 16MP reference image to the Image Conditioning node.  I'm attaching a comparison image.  The original image was a 1MP Anima image. The edit was to convert it to subtle realism.  I did three runs using a recently posted 20MP H3 Image edit workflow: no upscale, 4MP upscale, 20MP upscale.  The difference in gen times was 106s at no upscale, 133s at 20mp upscale.  If you can spot which image is which, feel certain the 20MP upscale at input is better aesthetically, and eat the 25% slower runtime difference to achieve that aesthetic, by all means upscale your ref. images accordingly.

[https://postimg.cc/H8srkJDv](https://postimg.cc/H8srkJDv)

document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xtbu7v/possible_flaw_in_h3_image_edit_default_workflow