crowmaki/ai-studio

Watch

Fork 0

Generation profile (#50) #104

Merged

crowmaki merged 3 commits from issue-50-generation-profile into main

2026-10-08 22:38:26 +00:00

crowmaki commented

2026-10-08 22:12:20 +00:00

Owner

Closes #50.

What changed

Generation profile (studio-core/src/generation_profile.rs): diffusion model file (any quantization), text encoder file, an ordered LoRA stack with strengths, Distilled/Full speed, and draft/final short sides. The shipped default reads its two files from the h3.draft template, so it can't drift from the workflows. It is Distilled at 768 / 1120 with no LoRAs.
Storage: migration V22 adds project_profile (a singleton) and project_profile_lora. A Project from before V22 gets the shipped profile. A new Project copies AppDefaults.generation_profile. An older defaults.json without that field loads with the shipped profile.
Edit: set_generation_profile replaces the whole profile as one undoable edit ("Change generation profile"). It trims file names and refuses blank files, a non-numeric LoRA strength, a short side under 32, and a draft larger than the final.
Pickers: comfyui_profile_choices lists /api/models/diffusion_models, text_encoders and loras. When ComfyUI is offline, it returns the reason instead of failing.
Preflight: preflight(client, Option<&GenerationProfile>). comfyui_preflight passes the open Project's profile. If a profile file is missing, the three H3 workflows are Not ready with a missing_profile_file reason. The image workflows don't use the profile, so the reason doesn't apply to them. A missing file is matched across \ and /. The render worker passes None for now, because it only runs image jobs. Checking a job's own profile before submission arrives with the draft Take queue (#51).
Staleness: the module doc says that the profile is never an input to staleness. I'll comment on #58 with the details.
LoRA switch: each LoRA can be switched off. It stays in the stack, dimmed, but is not applied, and preflight doesn't look for its file. The switch is a new Switch control in the UI kit (src/lib/ui/Switch.svelte, on Bits UI).
UI: the Generation profile settings page replaces the placeholder. The ComfyUI Section links a missing profile file to that page.

Screenshots

The pickers list what ComfyUI has. A LoRA that ComfyUI doesn't list is marked Missing:

Each LoRA has an on/off switch. Here a LoRA that ComfyUI lacks is switched on, so it is marked Missing and the H3 workflows are Not ready:

Switched off, it stays in the list without being applied, and ComfyUI is ready again:

When ComfyUI is offline, the page shows only the saved choices and the file pickers are disabled:

A missing profile file makes only the H3 workflows Not ready:

The screenshots use a stand-in ComfyUI that serves the shipped nodes and some made-up model files.

Testing

cargo test --workspace --all-features
bun test src scripts
bun run check
New tests:
- crates/studio-core/tests/generation_profile.rs covers defaults, app defaults, undo, validation, migration, preflight of missing profile files against the fake, and the pickers.
- src-tauri/tests/smoke.rs covers the adapter.
- src/lib/settings/profile.test.ts covers the picker items and the LoRA stack edits.
Checked by hand in the running app: picking a model, adding a LoRA, and a refused draft size, which shows the error and goes back to 768.

Review notes

The draft ≤ final rule (also a SQL CHECK) is my addition, because Promotion goes from the draft size up to the final size. Say if you'd rather allow it.
Rounding to ×32 and applying the speed mode to speed_modes both happen when the job is built (#51), not in this PR.

Closes #50. ## What changed - **Generation profile** (`studio-core/src/generation_profile.rs`): diffusion model file (any quantization), text encoder file, an ordered LoRA stack with strengths, Distilled/Full speed, and draft/final short sides. The shipped default reads its two files from the `h3.draft` template, so it can't drift from the workflows. It is Distilled at 768 / 1120 with no LoRAs. - **Storage**: migration V22 adds `project_profile` (a singleton) and `project_profile_lora`. A Project from before V22 gets the shipped profile. A new Project copies `AppDefaults.generation_profile`. An older `defaults.json` without that field loads with the shipped profile. - **Edit**: `set_generation_profile` replaces the whole profile as one undoable edit ("Change generation profile"). It trims file names and refuses blank files, a non-numeric LoRA strength, a short side under 32, and a draft larger than the final. - **Pickers**: `comfyui_profile_choices` lists `/api/models/diffusion_models`, `text_encoders` and `loras`. When ComfyUI is offline, it returns the reason instead of failing. - **Preflight**: `preflight(client, Option<&GenerationProfile>)`. `comfyui_preflight` passes the open Project's profile. If a profile file is missing, the three H3 workflows are Not ready with a `missing_profile_file` reason. The image workflows don't use the profile, so the reason doesn't apply to them. A missing file is matched across `\` and `/`. The render worker passes `None` for now, because it only runs image jobs. Checking a job's own profile before submission arrives with the draft Take queue (#51). - **Staleness**: the module doc says that the profile is never an input to staleness. I'll comment on #58 with the details. - **LoRA switch**: each LoRA can be switched off. It stays in the stack, dimmed, but is not applied, and preflight doesn't look for its file. The switch is a new `Switch` control in the UI kit (`src/lib/ui/Switch.svelte`, on Bits UI). - **UI**: the Generation profile settings page replaces the placeholder. The ComfyUI Section links a missing profile file to that page. ## Screenshots The pickers list what ComfyUI has. A LoRA that ComfyUI doesn't list is marked Missing: ![Generation profile page](https://git.crowmaki.me/attachments/0cf42030-0b98-41e8-a378-3b96dc425358) ![Diffusion model picker](https://git.crowmaki.me/attachments/22ab5919-8447-4350-ad75-a32b6d19ae3a) Each LoRA has an on/off switch. Here a LoRA that ComfyUI lacks is switched on, so it is marked Missing and the H3 workflows are Not ready: ![LoRA switched on](https://git.crowmaki.me/attachments/7d35ff39-feb0-4b4e-b5cf-9f293a593a77) Switched off, it stays in the list without being applied, and ComfyUI is ready again: ![LoRA switched off](https://git.crowmaki.me/attachments/23529c4b-f9f7-48b7-bc91-4e2117336bcd) When ComfyUI is offline, the page shows only the saved choices and the file pickers are disabled: ![Profile while ComfyUI is offline](https://git.crowmaki.me/attachments/9c8be37a-ab4b-48a3-bf5d-a278839b2ff4) A missing profile file makes only the H3 workflows Not ready: ![ComfyUI Section with a missing profile file](https://git.crowmaki.me/attachments/bcaefd2b-6cc4-4732-9473-b9619e92fc4d) The screenshots use a stand-in ComfyUI that serves the shipped nodes and some made-up model files. ## Testing - `cargo test --workspace --all-features` - `bun test src scripts` - `bun run check` - New tests: - `crates/studio-core/tests/generation_profile.rs` covers defaults, app defaults, undo, validation, migration, preflight of missing profile files against the fake, and the pickers. - `src-tauri/tests/smoke.rs` covers the adapter. - `src/lib/settings/profile.test.ts` covers the picker items and the LoRA stack edits. - Checked by hand in the running app: picking a model, adding a LoRA, and a refused draft size, which shows the error and goes back to 768. ## Review notes - The draft ≤ final rule (also a SQL CHECK) is my addition, because Promotion goes from the draft size up to the final size. Say if you'd rather allow it. - Rounding to ×32 and applying the speed mode to `speed_modes` both happen when the job is built (#51), not in this PR.

1-profile.png

61 KiB

2-picker.png

75 KiB

3-offline.png

68 KiB

4-comfyui.png

54 KiB

5-lora-on.png

68 KiB

6-lora-off.png

65 KiB

crowmaki added 2 commits

2026-10-08 22:12:20 +00:00

Generation profile (#50 ) 8011c02be9

A Project default Generation profile: diffusion model and text encoder
files, a LoRA stack with strengths, Distilled/Full speed and draft/final
short sides (768/1120). New Projects copy it from the app defaults, and
Projects from before it get the shipped H3 files.

The settings page picks files from the user's ComfyUI via /api/models/*.
Preflight checks the open Project's profile files on the H3 workflows.
The profile is never an input to staleness.

Address review of Generation profile (#50 ) 5266302159

Fix the run of spaces in the missing profile file message and pin it in
a test, note that V22's file names are frozen, and give the profile
view's refresh clearer names.

crowmaki referenced this pull request

2026-10-08 22:12:24 +00:00

Staleness reasons and acceptance #58

crowmaki added 1 commit

2026-10-08 22:37:14 +00:00

Switch each profile LoRA on or off (#50 ) 11c80b9a56

A LoRA switched off stays in the stack but is not applied, and preflight
no longer needs its file. The LoRA rows get a slider switch, a new
Switch control in the UI kit.

Also document loading the Forgejo token from .env for PR images.

crowmaki merged commit 500c84920d into main

2026-10-08 22:38:26 +00:00

crowmaki deleted branch issue-50-generation-profile

2026-10-08 22:38:27 +00:00

crowmaki referenced this pull request from a commit

2026-10-08 22:38:27 +00:00

Merge pull request 'Generation profile (#50)' (#104) from issue-50-generation-profile into main