ChitraDev

ChitraDev is a local-first FastAPI image generation app for Stable Diffusion 1.5. It can generate images from prompts alone, or use one or two FaceID reference photos through IP-Adapter FaceID identity guidance. Generated PNGs, ZIP downloads, reference photos, and model cache stay on your own machine.

ChitraDev is built to avoid hosted image-generation limits. There are no subscription quotas, remote credits, or repo-imposed creative category blocks in the app itself. Practical output is controlled by your local hardware, selected model behavior, environment settings, and the laws or consent requirements that apply to your use.

What ChitraDev Does

• Prompt-only image generation with Stable Diffusion 1.5.
• Optional single-person FaceID generation with one reference image.
• Optional two-person FaceID generation with separate Person 1 and Person 2 reference images.
• Left/right person placement controls for cleaner two-person identity separation.
• Previewable PNG outputs with direct image downloads and a ZIP download.
• Separate batch creator page for long prompt-only runs.
• Local FastAPI browser UI at http://127.0.0.1:8000.
• Public website lives in the separate ChitraDevWeb repository.

System Requirements

Recommended local setup:

Requirement	Recommended	Notes
OS	Windows 10/11	PowerShell scripts are included. Linux/macOS can adapt the same Python commands manually.
Python	3.11	The setup script creates a .venv using Python 3.11.
GPU	NVIDIA CUDA GPU	Defaults are tuned for a 4 GB GPU at 512x512, 20 steps, and 1 image.
CPU-only	Supported, but slow	CPU generation can work, but expect long generation times.
Disk	Several GB free	Stable Diffusion, IP-Adapter, and InsightFace assets cache in .cache/.
RAM	16 GB recommended	More memory helps model loading and larger batches.
Node.js	20 or newer	Needed only for the public website in frontend/.
Git	Latest stable	Needed to clone the repo and publish the website.
Network	Required on first model load	Use LOCAL_FILES_ONLY=true after assets are already cached locally.

Quick Local Start

1. Clone ChitraDev and open the folder:

git clone https://github.com/yuvrajraina/ChitraDev.git
cd ChitraDev

2. Optional: create .env from .env.example if you want custom local paths, default generation settings, or higher hardware caps:

Copy-Item .env.example .env

3. Run the setup script:

powershell -ExecutionPolicy Bypass -File .\scripts\setup.ps1

4. Activate the environment:

.\.venv\Scripts\Activate.ps1

5. Optional but recommended: preload the model stack before opening the UI:

powershell -ExecutionPolicy Bypass -File .\scripts\preload_models.ps1

6. Start ChitraDev:

python -m uvicorn app.main:app --host 127.0.0.1 --port 8000

7. Open the local app:

http://127.0.0.1:8000

If you see a torchvision::nms does not exist error, rerun the setup script. It force-reinstalls a matching CUDA build of torch, torchvision, and torchaudio.

How to Generate Images Locally

Prompt-only generation:

1. Open http://127.0.0.1:8000.
2. Leave both face upload fields empty.
3. Put the full subject, style, environment, lighting, and composition in the main prompt.
4. Click Generate images.

Single-person FaceID generation:

1. Upload a clear front-facing image in Person 1 face image.
2. Add the overall scene in Prompt.
3. Optionally describe clothing, pose, or action in Person 1 role and action.
4. Click Generate images.

Two-person FaceID generation:

1. Upload Person 1 face image and Person 2 face image.
2. Describe the overall scene in Prompt.
3. Add separate role/action text for each person if needed.
4. Use Interaction or relationship to describe what they are doing together.
5. Keep Person 1 and Person 2 on opposite sides in advanced settings.
6. Click Generate images.

Outputs are written to the ignored outputs/ folder and can be downloaded from the browser UI.

Batch generation:

1. Open http://127.0.0.1:8000/batch.
2. Paste one prompt that should be reused.
3. Set total outputs, chunk size, dimensions, steps, guidance, negative prompt, and optional start seed.
4. Click Start batch.

Batch jobs run in the background while the server stays open. Each job writes PNGs, batch-prompt.txt, and batch-results.zip into an ignored outputs/batch-.../ folder.

Generation Limits And Controls

ChitraDev does not add hosted image-generation limits, remote credits, subscription gates, or repo-level prompt category restrictions. It is meant to run locally, so you control the model, cache, outputs, and environment configuration.

There are still local safety rails for stability and hardware fit. These are configurable in .env:

Variable	Default	Purpose
DEFAULT_WIDTH	512	Default image width. Must be a multiple of 64.
DEFAULT_HEIGHT	512	Default image height. Must be a multiple of 64.
DEFAULT_STEPS	20	Default diffusion steps.
DEFAULT_NUM_IMAGES	1	Default images per request.
MAX_WIDTH	768	Maximum request width. Raise only if your GPU can handle it.
MAX_HEIGHT	768	Maximum request height. Raise only if your GPU can handle it.
MAX_STEPS	50	Maximum diffusion steps per request.
MAX_NUM_IMAGES	4	Maximum images per request.
MAX_GUIDANCE_SCALE	20	Maximum classifier-free guidance scale.
DEFAULT_BATCH_TOTAL_IMAGES	100	Default total outputs in the batch creator.
MAX_BATCH_TOTAL_IMAGES	500	Maximum outputs allowed for one batch job.
DEFAULT_BATCH_CHUNK_SIZE	1	Default images generated per batch chunk.
MAX_BATCH_CHUNK_SIZE	4	Maximum images generated per batch chunk.
MAX_FACE_IMAGE_MB	10	Maximum uploaded reference image size.

If you hit GPU memory issues, lower the image size, step count, or output count.

Configuration

Create .env from .env.example when you want to customize ChitraDev:

Copy-Item .env.example .env

Important settings:

• HOST and PORT control the local FastAPI address.
• OUTPUT_DIR controls where generated PNGs and ZIPs are written.
• CACHE_DIR controls where model assets are cached.
• SD_BASE_MODEL controls the Stable Diffusion model source.
• IP_ADAPTER_SOURCE and IP_ADAPTER_WEIGHT control FaceID adapter assets.
• HF_TOKEN can be used for gated Hugging Face assets.
• LOCAL_FILES_ONLY=true keeps model loading offline after assets are cached.
• DEFAULT_NEGATIVE_PROMPT sets the default negative prompt in the UI.

Health Check

curl http://127.0.0.1:8000/health

The response reports whether the service is alive and whether the model runtime has loaded.

Tests

The included tests cover settings loading, output saving and ZIP creation, prompt composition, and the web API with a stubbed inference service.

python -m unittest discover -s tests -v

Host The Website On Netlify

The ChitraDev app repo is the local FastAPI generator. Host only the separate website repo, https://github.com/yuvrajraina/ChitraDevWeb, on Netlify.

The frontend includes netlify.toml with the expected Vite settings:

Build command: npm run build
Publish directory: dist
Node version: 20

After Netlify gives you a live URL, update the frontend SEO URL values if needed:

• frontend/src/config/site.ts
• frontend/index.html
• frontend/public/robots.txt
• frontend/public/sitemap.xml