Document Qwen prompt alignment

Add the first-frame Qwen3-VL prompt-alignment workflow and mark its documentation complete.

README.md

@@ -48,6 +48,12 @@ world state through camera motion, occlusion, and illumination changes.
 This repository currently releases only the validated ReMind 5B weights. It
 does not contain ReMind 1.3B weights.
 
+Release checklist:
+
+- [x] Release ReMind-5B checkpoints
+- [x] Document Qwen3-VL prompt alignment for custom inputs
+- [ ] Release ReMind-1.3B checkpoints
+
 ## Files
 
 | File | Size | Purpose |
@@ -102,6 +108,46 @@ three-latent-frame chunks, and the shifted four-step schedule
 `[1000, 938, 833, 625]`. Camera examples use the bundled pair-fixed GT camera
 trajectories.
 
+## Qwen3-VL prompt alignment for custom inputs
+
+For a new first frame and an ordinary user request, we recommend using a local
+`Qwen/Qwen3-VL-8B-Instruct` call before ReMind inference. Give Qwen only the
+initial image and the user's text; do not provide future frames, evaluation
+criteria, or a hand-written target caption. Use greedy decoding and ask for a
+JSON object containing:
+
+- `scene_caption`: an English, content-only description of the initial state,
+  physical progression, and requested final state;
+- `control_requests`: any explicit `camera`, `occluder`, or `light` request,
+  with an exact supporting substring from the user text; and
+- `clarification_questions`: unresolved intent or conflicting controls.
+
+Preserve actions, negations, identities, directions, and stated end goals.
+Keep camera motion and reversible visibility events out of `scene_caption`:
+camera geometry belongs in a bundled trajectory, while occluder/light timing
+belongs in the preset's `control` block. Never ask Qwen to invent camera
+matrices or replace a bundled pair-fixed trajectory with an analytic path.
+
+Copy the accepted `scene_caption` into the closest preset, configure at most
+one supported control, and validate the resolved seven-chunk schedule before
+loading weights:
+
+```bash
+python inference.py --preset /path/to/custom_case.yaml --dry-run
+
+python inference.py \
+  --preset /path/to/custom_case.yaml \
+  --model-folder checkpoints/Wan2.2-TI2V-5B \
+  --base-checkpoint checkpoints/ReMind-5B/ReMind-5B.safetensors \
+  --ema-checkpoint checkpoints/ReMind-5B/ReMind-5b-dmd-ema.safetensors \
+  --output outputs/custom_case.mp4
+```
+
+Treat Qwen's rewrite and self-review as heuristic. Reject or manually review
+outputs that drop a negation, invent an event, mix control language into the
+scene caption, request multiple controls, or leave the intent ambiguous.
+Fixed project-page captions should be used verbatim rather than rewritten.
+
 ## Checkpoint lineage and validation
 
 - ReMind-5B generator: pair-fixed 480p checkpoint after an additional 2,000