06 — Tool Adapters
One concept, many engines. Build the 8-block stack once, then translate per the tool's parser and quirks.
REVE 2.0 (the contest tool) — deep notes
- Architecture: layout-first. Builds a structured, editable node graph (each element addressable with position/size/local description) then renders. Think "editing a document," not "re-rolling a prompt." → You can fix one hand/sign/face without disturbing lighting or the rest of the scene. This is the unfair advantage over MJ/Flux entrants.
- Strengths: best-in-class legible text (dedicated typography pass); native 4K (up to 4096²), print-ready (no upscale dice-roll); extreme prompt adherence (~9.5/10 — "one of the most obedient models tested" — honors horizon placement, spatial relationships, composition almost literally).
- Weaknesses: fine organic micro-detail (~7.7/10); dense multi-subject scenes, crowds, physical interactions (pouring liquid), and complex hands falter; crowds lack individual faces; faces/complex hands still trail GPT Image 2. "Studio-photograph" bias — it simulates the world as photographed in a controlled studio; natural/physics-heavy light can falter outside familiar photographic archetypes. Opinionated house aesthetic — push against it for "originality."
- Prompting: natural language, structured like a camera setup — open with a fixed spatial anchor (camera angle → environment → subject) THEN layer detail. Opposite of Midjourney keyword-soup. Vague prompts waste its biggest strength.
- Settings: native ~2048²; optional 4K upscale; supports 3:2, 16:9, etc.; <20s gen.
- Turn "enhance prompt" OFF for serious work — keep deterministic control; write the full prompt yourself.
- Negatives: no reliable negative field — use positive layout specification + targeted edits; state what you DO want, then edit out what slips in.
- References + conversational edits are first-class and lossless across iterations.
- Contest tactic: 1–2 hero subjects; make any crowd/hand the editable element you refine separately; frame the concept as a deliberate photographic setup.
Cross-tool quick table
| Tool | Prompt style | Negatives | Text | Notes |
|---|---|---|---|---|
| Reve 2.0 | natural, camera-setup order | ignored | excellent | layout-first edits; 4K; studio bias; 1–2 heroes |
| Midjourney v7 | keyword + --ar --no --style raw --stylize |
--no |
weak | strong aesthetic; --style raw + low --stylize for realism; 1:1 fewer distortions |
| Flux / Flux 2 | natural language, long descriptive | yes | very strong | rewards texture detail; great hands vs SDXL; open-weight |
| GPT-image / Image 2 | literal natural-language instructions | ignored→describe positively | highest | follows detailed briefs; best text integration |
| Imagen 3/4 | descriptive natural language | limited | strong | clean faces/hands/light/text; weak on small faces in busy scenes |
| Nano Banana 2 (Gemini) | conversational, multi-reference | ignored | strong | up to 14 refs; Google-Search-grounded; iterate across turns; see nano-banana-2 skill |
| Stable Diffusion / SDXL | tag + weighted negatives | strong | weak | ControlNet/inpaint/LoRA control; the neg-heavy workflow |
Translating one prompt across tools
- Reve / GPT-image / Imagen / Nano Banana: write the 8-block stack as flowing natural language; convert every "no X" to its positive opposite; drop weighted-negative syntax.
- Midjourney: compress to comma keywords; move forbidden items to
--no; add--style raw --stylize 50-150 --ar W:H; for realism avoid--stylize 600+. - SD/SDXL/Flux: keep the positive stack, ADD the universal weighted negative block (ch.01 §7), use ControlNet/OpenPose for pose/hand control, inpaint failures at 0.35–0.5 denoise.
- Realism floor that ports everywhere: real camera+lens+aperture, ONE named light + shadow direction, anti-plastic texture callouts, kill the blacklist words.
Routing by job
- Text-critical (logo, poster, signage): GPT-image or Ideogram or Reve. Never Midjourney for exact text.
- Skin/portrait realism: Reve, Seedream 4.5, Flux, or SD+inpaint.
- Wildlife/nature: Reve or Imagen or Flux with the NatGeo template (ch.02 §3).
- Multi-reference compositing / real-world grounding: Nano Banana 2.
- Maximum surgical control / consistent character: SD/Flux + ControlNet + LoRA.
- Hebrew text: none — composite in post (ch.03 §F).
Sources: curiousrefuge & decrypt & felloai & theplanettools & wavespeed & bestimageprompts (Reve 2.0 architecture/specs/prompting); venturebeat (Reve 1.0); artlist & z-image (Seedream); allaboutai (Imagen); nano-banana-2 skill (Gemini). Tool layer drifts monthly — re-verify before relying.