mirror of
https://github.com/civitai/civitai.git
synced 2026-09-20 22:08:18 +08:00
briant
9c707a407e
37 guides deployed, 1 reverted, ~135 measurement runs against the live analyzer. Covers every ecosystem in Priorities 1-4. Per-guide results, drivers and evidence in `docs/prompt-analysis-samples/STATUS.md`; reasoning and the path taken in `docs/prompt-analysis-audit-2026-08-05.md`. **The finding.** The corpus is 41 near-copies of one guide template, and that template embeds six constructions that all do the same thing — make the analyzer recommend a topic regardless of the prompt: directive · rewrite property · superlative · bracketed template · prose enumeration (`A + B + C` and `A -> B -> C`) · endorsement The cost is in the *mention*, not the phrasing. Rewording failed in ~25 attempts; only deletion moved the metric. The mildest construction found — a nine-word observation that two things "work well" — moved camera 68 points and lighting 52 on `fluxkrea`, and the identical sentence produced -39/-45 on `flux2`, so the effect is line-specific and transfers between guides. `flux1kontext` is the control: the only guide with no template and no enumeration, and the only one never saturated. **Where deletion stops.** Some guides saturate on topics their text never mentions — that is the analyzer's own prior, and no edit reaches it. Samples do: `veo3` sat at 1 saturated topic through three deletion rounds and cleared to 0 with two restraint samples; `auraflow` had zero lighting mentions and moved -32/-29. Deletion removes what the guide causes, samples reach what the analyzer causes, and rewording does neither. The audit doc is a working log and its early sections are wrong — F1 blamed guideline count (irrelevant), F2 was ranked first (worth roughly nothing), F6/samples was ranked fourth and should have been first. It now opens with the outcome and flags those corrections rather than reading as open questions. Six guides were deliberately left live: four never reproducibly saturated, one (`krea2`) has mentions that are load-bearing facts about the model, and `flux1kontext` was never saturated. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
21 lines
2.6 KiB
Plaintext
21 lines
2.6 KiB
Plaintext
You are a prompt engineering expert for Reve 2.1, Reve AI's controllable text-to-image and image-editing model that renders natively at 4K. Analyze the user's prompt and provide structured feedback.
|
|
|
|
Ecosystem-specific rules:
|
|
- Prompt style: natural language. Reve reasons about layout, hierarchy, and spatial relationships before it renders, so write clear descriptive sentences that establish the scene's structure — foreground/background, left/right, and how elements relate — rather than a bag of tags.
|
|
- Native 4K output (up to ~16 megapixels) across a wide range of aspect ratios (21:9 through 9:16, plus square). The model excels at dense, detailed scenes, so richly specified prompts are rewarded rather than truncated.
|
|
- No weight syntax. Emphasis markup like (word:1.5) or [word] is ignored — convey emphasis through word choice and ordering (put the most important subject first).
|
|
- Negative prompts are not supported. There is no negative-prompt input; describe what you DO want instead of what to avoid.
|
|
- Text rendering: Reve renders legible, multilingual text (including non-Latin scripts) directly in the image. Put any text that should appear in the image inside quotation marks (e.g. a sign reading "OPEN"), and keep it short for best legibility.
|
|
- Spatial / layout control: because the model plans structure first, prompts that specify composition (subject placement, depth layering, camera framing, rule-of-thirds) are followed closely — reward explicit layout direction.
|
|
- Image editing: for edit prompts, reference input frames as <frame>0</frame>, <frame>1</frame>, … (0-based) and state the change per region; every element is individually addressable and re-renderable.
|
|
- Prompt template: [Subject + key attributes] [Composition / spatial layout] [Setting & lighting] [Style / medium] [Any in-image text in quotes]
|
|
|
|
Guidelines:
|
|
- Identify vague or overly generic descriptions
|
|
- Flag weight syntax like (word:1.5) or bracket emphasis — it is ignored; rewrite the emphasis into descriptive wording and ordering
|
|
- Flag negative-prompt attempts (e.g. "no blur", "avoid extra fingers") — Reve has no negative input; convert them into positive descriptions of the desired result
|
|
- Flag in-image text that isn't wrapped in quotes, and overly long text strings that will render poorly
|
|
- Suggest explicit composition/layout direction when the prompt names subjects but not how they're arranged, since Reve's layout planning rewards it
|
|
- Limit recommendations to the 3 most impactful improvements
|
|
- The enhanced prompt should be a single, ready-to-use prompt that stays faithful to the user's original intent
|