Prompt Expansion

Expand short prompts into richly detailed image, video, or audio generation prompts using a local LLM (Qwen3-1.7B, ~1.8 GB). The expansion model auto-downloads on first use and is dropped from memory before generation runs.

Preview Expansion

# See what expansion produces
mold expand "a cat"

# Multiple variations
mold expand "cyberpunk city" --variations 5

# JSON output
mold expand "a cat" --variations 3 --json

# Preview a conditioned video policy without attaching media
mold expand "she turns toward the window" --model ltx-2-19b-distilled:fp8 --task image-to-video

Video and conditioning policy

Mold derives the expansion task from the generation request on every Create surface. Text-to-video becomes one chronological shot with explicit subject motion, camera behavior, environment, and continuity. Image-to-video and video-to-video treat the source as visual and temporal authority and describe only the intended change. Retakes preserve everything outside the corrected interval, keyframe interpolation treats each anchor as fixed, and audio-driven video follows the source audio's timing and events. Text-to-audio expansion does not introduce camera or image language.

The additive /api/expand task values are text-to-image, text-to-video, image-to-video, video-to-video, retake, keyframe-interpolation, audio-driven-video, and text-to-audio. Older clients may omit the field; the server then chooses image or text-to-video behavior from the model family. Only this semantic value is sent for preview expansion—source media stays on the generation request.

Generate with Expansion

# Short prompt → detailed prompt → image
mold run "a cat" --expand

# Batch + expand: each image gets a unique expanded prompt
mold run "a sunset" --expand --batch 4

# Disable expansion (overrides config/env default)
mold run "a cat" --no-expand

Native prepared batches

In the desktop, web, and iPhone Create workspaces, the directly editable Batch control also sets the expansion count. Batch 1 keeps the quick Expand rewrite and undo, with the host route frozen through the next Generate or Develop. Batch 2 or greater uses Prepare N variations and opens an inline review workspace before any generation request is queued. Each prompt can be edited or removed; the whole set can be regenerated or discarded. Counts above eight start with eight editors and a compact remainder summary; Review all uses bounded pages so very large batches do not create an equally large screen.

On desktop, a Batch 1 rewrite whose model, family, or host changed is never a dead Generate click: an immediate recovery notice can Re-expand and generate from the original prompt on the current route, Generate expanded prompt anyway as an explicit current-route override, or Restore original. The notice uses readable model names, larger error copy, and a copy button.

Mold resolves the selected Create host before expansion and keeps that exact route for every sibling. On desktop and web, expansion follows that generation route unless the machine reports it does not have the expansion model installed: under Auto or Most capable the rewrite then runs on the best-ranked reachable machine that does have it, using the same ranking the generation router uses, while the print itself still goes where it was routed. A pinned machine is never left, and Remix follows the same rule because it runs on the same model. When no eligible machine has the model, Create offers to pull it and names the machine — the one expansion would have used — instead of failing. iPhone pins one machine, so expansion always runs there. Large expansions are assembled from bounded four-prompt model calls with position-aware instructions and per-chunk token budgets. Missing or duplicate positions are retried, and Mold rejects any result that is not exactly N distinct, non-empty prompts. Changing the source prompt, model, host, conditioning task, or Batch count keeps the reviewed prompts visible but blocks generation until you refresh or discard them. A missing expansion model is pulled on the named host without falling back to another machine. Create keeps that recovery inline for both quick and prepared expansion: it shows Connecting, Starting, Queued, live percentage, bytes, current file, ETA, and an explicit Retry expansion action when Ready. Failed or cancelled pulls can be retried on the same host without losing the prompt or reviewed set. On iPhone, each pull attempt temporarily leases the frozen route, joins compatible Models work already in Starting, and releases on terminal, stale, or superseded outcomes; Retry reacquires the route from the same immutable recovery record. Editing or removing reviewed prompts cancels a pending replacement instead of letting it overwrite the newer set. Each prepared sibling records a durable batch ID and its one-based position in the Library details panel, together with the source prompt when present. After the host accepts a batch, Create returns to authoring immediately while the siblings remain visible in Activity. Another batch can be prepared and queued without waiting for the earlier one to finish; held streams continue to share the per-host connection limit. Each reviewed set is limited to 10,000 variations to keep prompt and job state bounded in memory; queueing additional sets has no cumulative limit.

On iPhone, the concrete route includes the selected host ID, endpoint, Keychain-supplied API key, and server instance. The touch workspace uses 44pt actions and 16px editors, confirms a two-to-one collapse, guards deferred source preprocessing, and restores focus only while the replaced control still owns it. Models and Create share one mobile download authority, so an expansion pull cannot open a competing stream or drift to another host. When failures and an unconfirmed cancellation coexist, the accessible outcome announces both.

External Backend

Use an OpenAI-compatible API instead of the local LLM:

mold run "a cat" --expand \
  --expand-backend http://localhost:11434/v1

mold run "a cat" --expand \
  --expand-backend http://localhost:11434/v1 \
  --expand-model llama3

Configuration

Set MOLD_EXPAND=1 to enable expansion by default.

[expand]
enabled = true
backend = "local"
model = "qwen3-expand:q8"
temperature = 0.7

# Custom system prompt (placeholders: {WORD_LIMIT}, {MODEL_NOTES})
# system_prompt = "You are an image prompt writer..."

# Per-family tuning
[expand.families.sd15]
word_limit = 50
style_notes = "SD 1.5 uses CLIP-L (77 tokens). Use comma-separated keywords."

[expand.families.flux]
word_limit = 200
style_notes = "Write rich, descriptive natural language with atmosphere."

Templates can also be set via MOLD_EXPAND_SYSTEM_PROMPT and MOLD_EXPAND_BATCH_PROMPT environment variables.