← Blog
✎Post blogowy

Two soft cover plates: we changed the words and the size at once

A macro camera lens barrel resting on a dark steel bench, a cold teal reflection on its front glass, a warm amber accent along the barrel, in teal and steel.

The re-roll of two soft plates at 1536 x 832 with 40 steps, cfg 6.5 and new positive and negative prompts produced the approved sharp set, and the sources do not show which of the five changes did it.

Two cover plates came back soft across the whole frame. The re-roll rendered them larger, with more steps, a higher guidance value and new words for both the prompt and the negative prompt. The guide now records the re-roll settings as the fix. Five things changed at once, so the sources cannot say which one did the work.

The soft plates

Both soft plates used the same settings. The render was 1344 x 704, which is 946,176 pixels, with 34 steps and a guidance value of 6.0. Their prompts included "shallow depth of field" and "fog". The negative prompt was a list of text and graphic terms with no blur words.

One of them was a sealed crate with a lit glass panel. The other was two identical engine blocks on a bench under teal light. Both are in the first job file under their ids, self-describing-bake and int8-no-gain.

The plate style guide describes the failure as a prompt problem first. Its rule: "tack sharp focus on , crisp edges, fine surface detail", atmosphere in the background only, and a negative prompt that names blur and soft focus. It also says that 1344 x 704 is barely a downscale, so the softness survives it.

The two prompts side by side

The first render of the engine-block plate asked for "two identical machined metal engine blocks on a dark bench under matched cold teal lights, perfectly symmetrical, one faint amber accent, still and clinical, fog, shallow depth of field, photorealistic, empty space between them". The re-roll asked for "two identical machined aluminium engine blocks side by side on a steel bench, tack sharp focus on both blocks, crisp machined edges, fine surface detail, matched cold teal lights, one faint amber accent, light haze in the background only, photorealistic, slight three quarter angle, empty space between them".

The subject and the lighting words are close to the same. The focus, edge and haze words are not. The re-roll also dropped "perfectly symmetrical" and "still and clinical", and changed "metal" to "aluminium". Those are wording changes that we did not test one at a time.

The guide's sizing rule is to render larger than the output and downscale on composite. A 1536 x 832 render comes down to 1200 x 630 for the page. The guide credits that supersampling step with the sharpness.

What the re-roll changed

The re-roll used the same two ids at 1536 x 832, which is 1,277,952 pixels and 1.35 times the first render. It used 40 steps instead of 34, and a guidance value of 6.5 instead of 6.0. The positive prompt swapped "shallow depth of field" and "fog" for "tack sharp focus", "crisp edges", "fine surface detail" and "light haze in the background only". The negative prompt gained "blurry, soft focus, out of focus, heavy bokeh, motion blur, hazy foreground".

That is five changes: size, steps, guidance, the positive words and the negative words. The guide's wording names the words and the render size as the cause together. It does not separate them.

What we did not measure

The job files do not record seeds. So the re-roll and the first render are the same two ids and roughly the same subject, but we cannot say they share a seed. Without a shared seed, a difference in one plate could be the seed as easily as the words.

The sources do not include a sharpness number for the re-rolls. The guide describes the re-roll settings as the fix and gives no measurement. We report that as the guide's claim, not as a measurement we made here.

What failed

Two plates came back soft on the first render. We did not run the isolation test before changing five things. That is the failure in this piece: a fix that works and cannot be attributed.

What we still do not know

Whether the words alone would have fixed the plates at 1344 x 704. Whether the size alone would have. Whether the guidance change did anything. The guide's rule is a good one, and it may be right. This article does not show it.

How to prove this wrong

Pick one soft plate and fix its seed. Render it four ways: old words at the old size and steps, old words at the new size and steps, new words at the old size and steps, and new words at the new size and steps. Measure sharpness on the same region in each. If the old words at the new size are as sharp as the new words, the words were not the cause. If the new words at the old size stay soft, the size was the cause. The pattern will say which one did it.

Provenance

  • /home/mario/Assets/comfy/_raw/2026-09-20/jobs.json, lines 26 to 28 (self-describing-bake at 1344 x 704, 34 steps, cfg 6.0, "shallow depth of field, fog") and lines 30 to 32 (int8-no-gain, same settings).
  • /home/mario/Assets/comfy/_raw/2026-09-20/redo-sharp.json, lines 2 to 4 (self-describing-bake re-roll at 1536 x 832, 40 steps, cfg 6.5, new positive and negative prompts) and lines 5 to 7 (int8-no-gain re-roll).
  • /home/mario/iTools/design/plate-forge/STYLE.md, lines 22 to 29 (the sharpness rule) and line 45 (1344 x 704 is barely a downscale).

Komentarze

Brak komentarzy.

Zaloguj się aby skomentować.