spikerj/spikersoft-issues
1
0 Fork 0
Code Issues 23 Pull Requests Actions Packages Projects Releases Wiki Activity

EPIC: Art Studio — record and replay generation parameters (+ text+image concept) #842

New Issue
Closed
opened 2026-07-25 20:21:28 +00:00 by spikerj · 2 comments
spikerj commented 2026-07-25 20:21:28 +00:00
Owner
Copy link
Copy source

Art Studio has never been able to say how anything was made, and its concept art has never been reproducible.

The root cause

SDXLLightningBackend._text_to_image only builds a seeded torch.Generator when a seed param arrives. The Art Studio path never sends one — ArtPipeStageOrchestrator.BuildJobParamsAsync sends literally {"prompt": <asset.Prompt>} — so generation falls through to the ambient torch RNG and the seed that produced an image is gone the moment it is made. Steps (4), resolution (1024), guidance (1.5) and the empty negative prompt are silent Python defaults that are never written down either.

The worker's result dict is exactly {output_path, output_type, image_path}; images are saved with a bare image.save(path) — no PNG text chunk, no EXIF, no sidecar.

Consequence: POST {id}/restart/{stage} re-runs the same prompt and produces a different picture every time, and clicking an asset card shows name/type/status but nothing about the prompt, model or settings.

Scope

  1. Record what actually produced each stage run, per attempt.
  2. Always seed generation, so a run is reproducible at all.
  3. Show it when you click a card.
  4. Let a user reproduce a run exactly, or vary it deliberately.
  5. Let a user set a seed / negative prompt / steps up front.
  6. Open the text+image concept path (prompt + reference image), which did not exist.

Shaped so FLUX.2 drops in as a concept model without redesigning the contract: references are a list (FLUX.2 takes up to ten), strength is optional and backend-declared (FLUX.2 has none), and steps/guidance bounds come from per-method capability descriptors rather than a hardcoded list.

Sub-issues

See the linked issues.

Art Studio has never been able to say how anything was made, and its concept art has never been reproducible. ## The root cause `SDXLLightningBackend._text_to_image` only builds a seeded `torch.Generator` when a `seed` param arrives. The Art Studio path never sends one — `ArtPipeStageOrchestrator.BuildJobParamsAsync` sends literally `{"prompt": <asset.Prompt>}` — so generation falls through to the ambient torch RNG and **the seed that produced an image is gone the moment it is made**. Steps (4), resolution (1024), guidance (1.5) and the empty negative prompt are silent Python defaults that are never written down either. The worker's result dict is exactly `{output_path, output_type, image_path}`; images are saved with a bare `image.save(path)` — no PNG text chunk, no EXIF, no sidecar. Consequence: `POST {id}/restart/{stage}` re-runs the same prompt and produces a **different picture every time**, and clicking an asset card shows name/type/status but nothing about the prompt, model or settings. ## Scope 1. Record what actually produced each stage run, per attempt. 2. Always seed generation, so a run is reproducible at all. 3. Show it when you click a card. 4. Let a user reproduce a run exactly, or vary it deliberately. 5. Let a user set a seed / negative prompt / steps up front. 6. Open the **text+image** concept path (prompt + reference image), which did not exist. Shaped so FLUX.2 drops in as a concept model without redesigning the contract: references are a list (FLUX.2 takes up to ten), `strength` is optional and backend-declared (FLUX.2 has none), and steps/guidance bounds come from per-method capability descriptors rather than a hardcoded list. ## Sub-issues See the linked issues.
spikerj commented 2026-07-29 15:38:58 +00:00
Author
Owner
Copy link
Copy source

Epic status roll-up — audited every child against origin/master / origin/main. All seven shipped and merged (epic PR d0de275a / #480), but three carry real gaps, so the epic stays open until they land.

Child State
#843 record stage-run params ✅ closed
#844 artpipe seed + provenance ✅ closed
#848 UI: show + remake ✅ closed
#849 name storage keys ✅ closed (superseded by #853)
#845 carry overrides to worker ⚠️ worker clamps silently; GET options unfiltered by source type
#846 text+image concept ⚠️ restart drops the source type
#847 reproduce and vary ⚠️ dead Enabled switch; MethodKey accepted and ignored

The three things blocking this epic, in the order I'd fix them:

  1. #846 — restart silently reverts text+image to text. RestartFromStageCommandHandler.cs:390 calls ToStageMethod(option) without the source type, so ActionFor falls back to the text action and the reference images are ignored. One-argument fix, but it's the exact silent-wrong-result behaviour the epic set out to remove.
  2. #847 — ArtStudio:GenerationParams:Enabled is dead code. Declared and documented as "False rejects every override", never read by anything. Overrides are accepted in production today regardless of the flag. Either wire it in or delete the flag and its doc comment — right now the documentation asserts a safety property that doesn't exist.
  3. #845 — the worker's "drop with a warning" is silent. Clamping happens with no log line anywhere, so a clamped override is invisible to both the user and anyone reading worker logs.

Two smaller items also recorded on their children: #847's MethodKey reaches the command and is never read (so "reproduce on a different model" silently uses the same model), and #846's SupportsSourceType guard only covers the explicit selection loop — preset-merge, failover and darkroom bypass it.

Everything else in the epic verified clean, including the cross-repo contract: the backend mapper's KnownKeys match artpipe/src/artpipe/generation_params.py:150-221 key-for-key, pinned from both sides.

**Epic status roll-up** — audited every child against `origin/master` / `origin/main`. All seven shipped and merged (epic PR `d0de275a` / #480), but three carry real gaps, so the epic stays open until they land. | Child | State | |---|---| | #843 record stage-run params | ✅ closed | | #844 artpipe seed + provenance | ✅ closed | | #848 UI: show + remake | ✅ closed | | #849 name storage keys | ✅ closed (superseded by #853) | | #845 carry overrides to worker | ⚠️ worker clamps silently; `GET options` unfiltered by source type | | #846 text+image concept | ⚠️ **restart drops the source type** | | #847 reproduce and vary | ⚠️ dead `Enabled` switch; `MethodKey` accepted and ignored | **The three things blocking this epic**, in the order I'd fix them: 1. **#846 — restart silently reverts text+image to text.** `RestartFromStageCommandHandler.cs:390` calls `ToStageMethod(option)` without the source type, so `ActionFor` falls back to the text action and the reference images are ignored. One-argument fix, but it's the exact silent-wrong-result behaviour the epic set out to remove. 2. **#847 — `ArtStudio:GenerationParams:Enabled` is dead code.** Declared and documented as "False rejects every override", never read by anything. Overrides are accepted in production today regardless of the flag. Either wire it in or delete the flag and its doc comment — right now the documentation asserts a safety property that doesn't exist. 3. **#845 — the worker's "drop with a warning" is silent.** Clamping happens with no log line anywhere, so a clamped override is invisible to both the user and anyone reading worker logs. Two smaller items also recorded on their children: #847's `MethodKey` reaches the command and is never read (so "reproduce on a different model" silently uses the same model), and #846's `SupportsSourceType` guard only covers the explicit selection loop — preset-merge, failover and darkroom bypass it. Everything else in the epic verified clean, including the cross-repo contract: the backend mapper's `KnownKeys` match `artpipe/src/artpipe/generation_params.py:150-221` key-for-key, pinned from both sides.
spikerj commented 2026-08-06 00:20:35 +00:00
Author
Owner
Copy link
Copy source

Epic complete — all six scope items shipped across the three repos.

Backend (spikersoft-backend#480, "Closes #843, #845, #846, #847, #849"):

  • ArtAssetStageRun records the worker's generation_params per attempt (#843) — on the run, not the asset, because StageMethodProvenance is upsert-by-stage and would destroy earlier attempts
  • RestartFromStageCommand.GenerationParamOverrides carried to the worker as a field separate from static ArtAssetStageMethod.Params, with prompt/negative_prompt rejected rather than clamped so the restart path can't launder a moderated prompt (#845)
  • ArtAssetSourceType.TextAndImage + the SubmitArtAssetValidator ObjectId bug that was failing every image-upload submission in production post-MinIO (#846)
  • ReproduceStageRunId / ArtAssetReproduceMode / ReproducedFromStageRunId for exact-reproduce vs. deliberate-vary (#847)

artpipe — generation is always seeded now: sdxl_lightning.py:380,436, sdxl_turbo.py:97, flux_schnell.py, text3d.py all build a torch.Generator(...).manual_seed(seed), with the CUDA-vs-MPS/CPU non-reproducibility caveat documented at sdxl_lightning.py:470.

Angular — libraries/features/art-studio/src/lib/art-generation-details/ (component + dialog) surfaces the recorded parameters when you click a card, with a per-attempt Reproduce action (art-generation-details.component.html:251-256) and isReproducible() in art-studio.models.ts refusing to call a run reproducible without a recorded seed (seed 0 handled correctly).

— verified and closed by Opus 5 Agent during the 2026-08-06 tracker sweep.

**Epic complete** — all six scope items shipped across the three repos. **Backend** (spikersoft-backend#480, "Closes #843, #845, #846, #847, #849"): - `ArtAssetStageRun` records the worker's `generation_params` per attempt (#843) — on the *run*, not the asset, because `StageMethodProvenance` is upsert-by-stage and would destroy earlier attempts - `RestartFromStageCommand.GenerationParamOverrides` carried to the worker as a field separate from static `ArtAssetStageMethod.Params`, with `prompt`/`negative_prompt` **rejected rather than clamped** so the restart path can't launder a moderated prompt (#845) - `ArtAssetSourceType.TextAndImage` + the `SubmitArtAssetValidator` ObjectId bug that was failing **every** image-upload submission in production post-MinIO (#846) - `ReproduceStageRunId` / `ArtAssetReproduceMode` / `ReproducedFromStageRunId` for exact-reproduce vs. deliberate-vary (#847) **artpipe** — generation is always seeded now: `sdxl_lightning.py:380,436`, `sdxl_turbo.py:97`, `flux_schnell.py`, `text3d.py` all build a `torch.Generator(...).manual_seed(seed)`, with the CUDA-vs-MPS/CPU non-reproducibility caveat documented at `sdxl_lightning.py:470`. **Angular** — `libraries/features/art-studio/src/lib/art-generation-details/` (component + dialog) surfaces the recorded parameters when you click a card, with a per-attempt **Reproduce** action (`art-generation-details.component.html:251-256`) and `isReproducible()` in `art-studio.models.ts` refusing to call a run reproducible without a recorded seed (seed `0` handled correctly). — verified and closed by Opus 5 Agent during the 2026-08-06 tracker sweep.
Sign in to join this conversation.
No Branch/Tag Specified
Labels
Clear labels
agentic
Actionable surfaced by the agentic architecture & security review
 bug
Something is not working
 duplicate
This issue or pull request already exists
 enhancement
New feature
 epic
Cross-repo epic — children live in per-repo trackers
 help wanted
Need some help
 invalid
Something is wrong
 question
More information is needed
 sonarqube
Code quality issue surfaced by SonarQube (bug / code smell)
 wontfix
This won't be fixed
No labels
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
spikerj (Joseph Spiker)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: spikerj/spikersoft-issues#842