Build a disciplined image prompt iteration loop in the CLI
Change one variable per run, record model and ratio, and control credit spend.
OfflineCreator's current Studio catalog lists FLUX Schnell at 1 credit, FLUX 1.1 Pro Ultra at 8 credits, and Recraft V3 at 5 credits per generation. The catalog positions them respectively for fast iteration, photoreal detail, and brand graphics. Those labels help route a proof; they are not a quality ranking or evidence that a particular prompt will pass. The same first-party page says web, MCP, CLI, and API generations share the account credit balance and published per-model costs.
Set a run cap before the baseline. For example, one FLUX Schnell baseline plus two one-variable revisions has a published catalog cost of 6 credits at the cited 1-credit rate. Stop after those three runs and review the contact sheet; do not silently expand the batch because the last result is almost right. Re-run `models` and `balance` immediately before spending because catalog prices and account balances are mutable. OfflineCreator says selected costs are visible before generation and reserved credits return when a provider job fails, but a successful yet unusable output is still an iteration cost.
- Fast proof
- FLUX Schnell · 1 creditA three-run baseline-and-two-variants loop totals 3 credits at the currently published rate.
- Photoreal-detail route
- FLUX 1.1 Pro Ultra · 8 creditsUse the live model listing as the final cost authority before each approved run.
- Brand-graphics route
- Recraft V3 · 5 creditsThe catalog label describes intended fit, not a guaranteed result.
Freeze a baseline before changing the prompt
Begin with a written test card, not a sequence of improvised commands. Record the deliverable, selected model, aspect ratio, exact prompt, expected visual facts, and one pass condition. A useful baseline might require a matte green ceramic bottle on pale limestone, a low three-quarter camera, soft side light, an uncluttered left third, no lettering, and a 16:9 frame. The pass condition should name observable evidence: the bottle silhouette is complete, the left third remains quiet, and no pseudo-text appears.
Use the version-pinned command surface: `npx @offlinecreator/mcp@0.1.2 models`, then `npx @offlinecreator/mcp@0.1.2 balance`, then `npx @offlinecreator/mcp@0.1.2 generate --model "$MODEL" --prompt "$PROMPT" --aspect 16:9 --wait`. The published CLI requires a model and prompt, accepts `--aspect` or `--aspectRatio`, recognizes 1:1, 16:9, and 9:16, and can wait for a text-to-image result. Read the model ID and balance at runtime rather than copying a display name or remembered account value.
Change exactly one prompt variable per paid run
Copy the accepted baseline into a new shell variable and alter one clause. If the bottle is too centered, change only “centered bottle” to “bottle anchored in the right third.” Keep the model, ratio, lighting, materials, camera, exclusions, and pass condition unchanged. Name the run `v02-position-right`, save the exact prompt beside the returned generation identifier, and compare it with `v01-baseline`. A later reviewer can then connect a visible difference to one intended change instead of guessing among a rewritten prompt's many changes.
Do not claim deterministic reproduction from this CLI loop. The current 0.1.2 generate command forwards a model ID, prompt, and recognized aspect ratio; its published implementation exposes no seed flag. An upstream model API may offer a seed, but that does not create an undocumented Studio CLI control. The disciplined unit here is a controlled comparison: one changed instruction, the same review matrix, and an honest record of the output. If unrelated details drift, record that drift rather than crediting the changed clause with every difference.
Keep prompt history separate from account history
Maintain a local manifest for the experiment even though Studio keeps account records. Use one row per attempt with: local run label, timestamp, model ID, displayed credit cost, aspect ratio, exact prompt, the single changed clause, returned generation ID, output filename, pass or fail, and a short defect note. Studio's privacy policy identifies prompts, selected model, aspect ratio, workflow, and generated media as generation content, and says generation history and media are retained until the user deletes them or closes the account, subject to short backup windows.
That retention statement does not prove that the thin CLI can export complete prompt history. The package documents `list_generations` as an MCP tool for recent jobs, while the CLI's published commands are models, balance or credits, top-ups, and generate. Do not invent a `history` command. Treat the local manifest as the reproducible editorial record, the Studio account as the service record, and any recent-jobs tool as a reconciliation aid. Never put API keys, client secrets, private filenames, or undisclosed campaign facts in either the prompt or a shared manifest.
- Input identity
- model · ratio · exact promptCopy values from the submitted command and current model listing.
- Controlled delta
- one changed clauseState both the old and new wording so the comparison remains auditable.
- Observed result
- generation ID · output · pass/failDescribe visible defects without claiming the prompt alone caused every variation.
Choose the next guide from the failed variable
Return to the creative-workflow directory when the model or deliverable remains undecided. Move to the camera-motion guide only after an approved still needs motion language. Use the video-iteration guide when the expensive variable is a video prompt, wait, or cancellation path rather than a still-image clause. These links represent different production decisions; they are not alternate pages for the same image-prompt query.
Stay on this page while the unresolved question can be isolated inside one text-to-image run: subject placement, material wording, lighting direction, amount of negative space, or an exclusion. Escalate to manual compositing or another supported workflow when a requirement needs an input the current CLI does not expose. Repeating an undocumented flag or rewriting every clause does not make the test more controlled.
Keep the evidence boundary around the iteration record
This page owns the narrow task of iterating text-to-image prompts through the current OfflineCreator CLI while holding model and ratio constant, changing one clause, recording prompt and account evidence, and enforcing a credit cap. It does not promise deterministic images, a hidden seed, automatic prompt versioning, complete history export from the CLI, a particular number of successful attempts, or better output from a higher-cost model. Current costs and command behavior must be rechecked before use.
The mandatory recent-evidence run was degraded. It returned 68 items across GitHub, web grounding, Hacker News, Instagram, Reddit, TikTok, and YouTube. Two Reddit records were marginal context about same-prompt video comparison or prompt adherence, but neither documented this image-CLI workflow; the other records were irrelevant, duplicate, promotional, stale, or otherwise unusable. Reddit was partial, X was not configured, and Polymarket had no source-status record. Those outcomes do not show that affected communities were quiet. No community item supports OfflineCreator product behavior or practitioner consensus here. The page remains an unpublished, unreviewed researched draft grounded in current first-party product material and an explicit test procedure.