aladac/tensors
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update

Browse Source
This commit is contained in:
Adam Ladachowski
2026-02-12 16:40:49 +00:00
parent 795cc84122
commit 0d278faab2
3 changed files with 243 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
PLAN.md
+147
View File
@@ -0,0 +1,147 @@
# Plan: SD Image Generation (CLI + Web Gallery)
Add `tsr gen` CLI command and `tsr gallery` web UI for generating images using diffusers + PyTorch (ROCm) directly from local safetensor checkpoints. Gallery is mobile-first, generation-only — no search, no model library.
## Stack
- **diffusers** — load safetensor checkpoints via `from_single_file()`, LoRA via `load_lora_weights()`, all schedulers built-in
- **torch** (ROCm) — assumed pre-installed with ROCm support (`torch.device("cuda")` works on ROCm via HIP)
- **transformers** — CLIP text encoders (diffusers dependency)
- **accelerate** — device placement
- **safetensors** — already a project dependency
## Phase 1: Generation Engine Module
### Description
New `tensors/generate.py` module — wraps diffusers pipeline for txt2img from local safetensor files. Handles checkpoint loading, LoRA application, scheduler selection, and generation.
### Steps
#### Step 1.1: Create generate.py with pipeline management
- **Objective**: Load safetensor checkpoints into diffusers pipeline, generate images
- **Files**: `tensors/generate.py`
- **Dependencies**: None
- **Implementation**:
- `ImageGenerator` class
- `load_checkpoint(path)` — `StableDiffusionPipeline.from_single_file()` or `StableDiffusionXLPipeline.from_single_file()` for SDXL safetensors. Auto-detect SD1.5 vs SDXL from metadata. Move to ROCm device.
- `load_lora(path, strength)` — `pipe.load_lora_weights()`, support multiple LoRAs with `pipe.fuse_lora(lora_scale=strength)`
- `set_scheduler(name)` — map name strings to diffusers schedulers: EulerDiscreteScheduler, EulerAncestralDiscreteScheduler, DPMSolverMultistepScheduler, DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler, etc.
- `generate()` — accepts prompt, negative_prompt, steps, cfg_scale, width, height, seed, batch_size. Returns list of PIL Images.
- Keep pipeline loaded between calls (model stays in VRAM)
- `unload()` — free VRAM
#### Step 1.2: Add config entries for generation defaults
- **Objective**: Configurable default generation params and model paths in config.toml
- **Files**: `tensors/config.py`
- **Dependencies**: Step 1.1
- **Implementation**:
- Add `[generate]` section: `models_dir` (default `~/models`), `lora_dir`, `output_dir` (default `~/.local/share/tensors/gallery/`), `default_steps`, `default_cfg`, `default_sampler`, `default_scheduler`, `default_width`, `default_height`
- Enum or list of available schedulers with friendly names
## Phase 2: CLI Generate Command
### Description
`tsr gen` command that loads a checkpoint, generates images, saves to gallery directory with metadata sidecar JSON.
### Steps
#### Step 2.1: Implement `tsr gen` command
- **Objective**: CLI command with all generation parameters as options
- **Files**: `tensors/cli.py`
- **Dependencies**: Phase 1
- **Implementation**:
- `tsr gen "prompt text"` — positional prompt argument
- Options: `--model/-m` (path or name in models_dir), `--negative/-n`, `--steps/-s`, `--cfg/-c`, `--width/-W`, `--height/-H`, `--sampler`, `--scheduler`, `--seed`, `--lora` (repeatable, format `name:strength`), `--batch/-b`, `--output/-o` (override output dir)
- Rich progress: model loading spinner, then generation progress (diffusers callback for step progress)
- Save output as `{timestamp}_{seed}.png` in gallery dir
- Save sidecar `{timestamp}_{seed}.json` with all generation params + model name + time elapsed
- Display: filename, resolution, seed, time elapsed
- `--json` flag for machine-readable output
#### Step 2.2: Add `tsr gen-ls` subcommand
- **Objective**: List available models, LoRAs, and schedulers
- **Files**: `tensors/cli.py`
- **Dependencies**: Phase 1
- **Implementation**:
- Scan models_dir for `.safetensors` files
- Scan lora_dir for LoRA files
- List available schedulers
- Rich table output, `--json` flag
## Phase 3: Web Gallery UI
### Description
`tsr gallery` serves a mobile-first web app for generating images and browsing results. Single HTML file, no build tools. Dark theme.
### Steps
#### Step 3.1: Create gallery API server
- **Objective**: FastAPI app that runs generation and serves the gallery
- **Files**: `tensors/gallery.py`
- **Dependencies**: Phase 1, Phase 2
- **Implementation**:
- Holds a single `ImageGenerator` instance (lazy-loaded on first generate)
- `POST /api/generate` — accepts generation params JSON, runs pipeline, saves to gallery dir, returns image URL + metadata. Checkpoint loaded/swapped as needed.
- `GET /api/images` — list gallery images (paginated, newest first), reads sidecar JSONs for metadata
- `GET /api/images/{filename}` — serve image file
- `DELETE /api/images/{filename}` — delete image + sidecar
- `GET /api/models` — list available checkpoints in models_dir
- `GET /api/loras` — list available LoRAs
- `GET /api/schedulers` — list available scheduler names
- `GET /api/config` — current default generation params
- `GET /api/status` — is model loaded, which one, VRAM usage
- Static file serving for the frontend
- Add `fastapi`, `uvicorn` as optional dependencies (`[project.optional-dependencies] gallery = [...]`)
#### Step 3.2: Build mobile-first gallery frontend
- **Objective**: Single-page responsive UI for generation + browsing
- **Files**: `tensors/static/index.html`
- **Dependencies**: Step 3.1
- **Implementation**:
- **Generate panel** (top on mobile, sidebar on desktop):
- Model selector dropdown (populated from `/api/models`)
- Prompt textarea, negative prompt textarea
- Collapsible "Advanced" section: steps, cfg, sampler dropdown, scheduler dropdown, width, height, seed, LoRA selector with strength slider
- Generate button with loading state + step progress
- Dropdowns populated from API on load
- **Gallery grid** (below/main area):
- Masonry or uniform grid of generated images, newest first
- Tap/click to view full size with metadata overlay (prompt, params, seed)
- Swipe between images on mobile
- Delete button on detail view
- Infinite scroll / load more
- **Design**:
- Dark theme
- CSS grid/flexbox, no framework
- Touch-friendly (large tap targets, no hover-dependent UI)
- `<meta name="viewport">` for mobile
- Single HTML file with inline CSS/JS (no build step)
#### Step 3.3: Add `tsr gallery` CLI command
- **Objective**: Launch the gallery web server from CLI
- **Files**: `tensors/cli.py`
- **Dependencies**: Step 3.1, Step 3.2
- **Implementation**:
- `tsr gallery` — starts uvicorn on `0.0.0.0:7860`
- Options: `--port/-p`, `--host`, `--model/-m` (pre-load a checkpoint)
- Auto-open browser with `--open` flag
## Phase 4: Tests
### Steps
#### Step 4.1: Test generate.py
- **Files**: `tests/test_generate.py`
- **Dependencies**: Phase 1
- **Implementation**:
- Mock torch/diffusers (don't require GPU in CI)
- Test scheduler mapping, parameter validation, config loading
- Test checkpoint type detection (SD1.5 vs SDXL)
#### Step 4.2: Test gallery API
- **Files**: `tests/test_gallery.py`
- **Dependencies**: Phase 3
- **Implementation**:
- Use FastAPI TestClient
- Mock ImageGenerator
- Test image listing, deletion, model/lora listing
TODO.md
+19 -8
View File
@@ -1,8 +1,25 @@
# TODO
## Web UI
## SD Image Generation
Add a web interface started from the CLI (`tsr serve` or `tsr ui`) to:
### Phase 1: Generation Engine
- [ ] Step 1.1: Create `tensors/generate.py` — ImageGenerator class (diffusers pipeline, checkpoint loading, LoRA, schedulers)
- [ ] Step 1.2: Add `[generate]` config section (models_dir, lora_dir, output_dir, defaults)
### Phase 2: CLI Generate Command
- [ ] Step 2.1: `tsr gen` command (prompt, model, negative, steps, cfg, sampler, scheduler, seed, lora, resolution)
- [ ] Step 2.2: `tsr gen-ls` command (list models, LoRAs, schedulers)
### Phase 3: Web Gallery
- [ ] Step 3.1: `tensors/gallery.py` — FastAPI server (generate, images, models, loras, schedulers endpoints)
- [ ] Step 3.2: `tensors/static/index.html` — mobile-first dark gallery UI (generate panel + image grid)
- [ ] Step 3.3: `tsr gallery` CLI command (launch server)
### Phase 4: Tests
- [ ] Step 4.1: `tests/test_generate.py` (mocked diffusers, scheduler mapping, config)
- [ ] Step 4.2: `tests/test_gallery.py` (FastAPI TestClient, mocked generator)
## Web UI (Future)
### Model Library
- [ ] Browse downloaded models in `~/.local/share/tensors/models/`
@@ -19,9 +36,3 @@ Add a web interface started from the CLI (`tsr serve` or `tsr ui`) to:
- [ ] View model details and versions
- [ ] One-click download to appropriate directory
- [ ] Show download progress
### Technical
- [ ] Use FastAPI + htmx or similar lightweight stack
- [ ] SQLite for local model index/cache
- [ ] Watch filesystem for new models
- [ ] Configurable port (`tsr serve --port 8080`)
models.md
+77
View File
@@ -0,0 +1,77 @@
# Model Inventory
Location: `/models/`
## Checkpoints
### SD 1.5 (512x512)
| Model | Size | GGUF | Tensors | Notes |
|-------|------|------|---------|-------|
| dreamshaper_8 | 2.0G | 1.7G | 1,131 | |
| epicrealism_naturalSinRC1VAE | 2.0G | 1.7G | 1,133 | Includes VAE |
| hassakuSD15_v13 | 2.0G | 1.7G | 1,131 | |
### SDXL / Pony / Illustrious (1024x1024)
| Model | Size | GGUF | Tensors | Notes |
|-------|------|------|---------|-------|
| cyberrealisticPony_v160 | 6.5G | 3.9G | 2,515 | Author: Cyberdelia |
| juggernautXL_ragnarokBy | 6.6G | 4.0G | 2,516 | |
| obsessiveCompulsive_v20 | 6.5G | 3.9G | 2,515 | |
| ponyDiffusionV6XL_v6StartWithThisOne | 6.5G | 3.9G | 2,515 | arch: `stable-diffusion-xl-v1-base` |
| ponyRealism_V22 | 6.6G | 4.0G | 2,515 | Merge: PonyRealism v2.1 + Volendir Cinematic v1.1R |
| realismIllustriousBy_v50FP16 | 6.5G | 3.9G | 2,515 | |
| spicyRealismNSFWMix_v30 | 6.5G | 3.9G | 2,515 | Triple merge |
| waiIllustriousSDXL_v160 | 6.5G | 3.9G | 2,515 | |
**Total checkpoints:** 11 models, ~95G (safetensors + GGUF)
## LoRAs
### SDXL / Illustrious
| LoRA | Size | GGUF | Dim | Alpha | Clip Skip | Title |
|------|------|------|-----|-------|-----------|-------|
| 70s_VPMS_V1-E20 | 218M | - | 32 | 32 | 2 | 70s Vintage Porn Magazine Style |
| Bimbo_Bomb_Girls_Pit_Style | 218M | 116M | 32 | 16 | 1 | Bimbo Bomb Girls Pit Style |
| Candy_Jab_Comix | 218M | 116M | 32 | 16 | 1 | Candy Jab Comix |
| Nellie_Jab_Comix-000009 | 218M | - | 32 | 16 | 1 | Nellie Jab Comix |
| RealisticAnimeIXL_v2 | 218M | 116M | 32 | 16 | 1 | RealisticAnimeIXL |
| Western_art_style (Melkor Mancin / Rizdraws) | 218M | - | 32 | 16 | 1 | Combined western art style |
| spumcostyle | 218M | 116M | 32 | 16 | 1 | spumcostyle |
| vitpitillust | 218M | 116M | 32 | 16 | - | vitpitillust |
### SD 1.5
| LoRA | Size | GGUF | Dim | Alpha | Clip Skip | Title |
|------|------|------|-----|-------|-----------|-------|
| BimboOne | 144M | 82M | 128 | 128 | 2 | BimboOne |
| Calm [MockAI - v1.0] | 37M | - | - | - | - | - |
| bimbo-fc-1.6a | 144M | - | - | - | - | - |
| bimbostyleTwo | 144M | 82M | 128 | 128 | 2 | bimbostyleTwo |
**Total LoRAs:** 12 models, ~3.4G (safetensors + GGUF)
## Character Presets
YAML files in `/models/characters/` with trigger words, positive/negative prompts per LoRA.
| LoRA | Presets |
|------|---------|
| 70s_VPMS | generic |
| BimboOne | candy_charms |
| Bimbo_Bomb_Girls_Pit_Style | blossom, bubbles, butter, generic |
| Calm_MockAI | generic |
| Candy_Jab_Comix | candy |
| Nellie_Jab_Comix | nellie |
| RealisticAnimeIXL | generic |
| Western_Melkor_Mancin | generic |
| bimbo-fc | generic |
| bimbostyleTwo | generic |
| spumcostyle | generic |
| vitpitillust | bimbo |
## Outputs
`/models/outputs/` — 167 generated images, 141M total