Skip to content

AbstractVision documentation

This folder contains the user-facing documentation for abstractvision.

Start here (new users)

1) Project overview + quickstart
2) Getting started (shell and interactive t2i/i2i; then Diffusers, MLX-Gen, GGUF, OpenAI-compatible HTTP, Playground) 3) Architecture (how the pieces fit together)

Quick reference

AbstractFramework ecosystem

AbstractVision is part of the AbstractFramework ecosystem and is designed to compose with:

Current implementation status (as shipped)

Public API surface: VisionManager exposes: - generate_image (text_to_image), edit_image (image_to_image) - generate_image_batch, edit_image_batch (repeatable orchestration with explicit seed planning) - generate_video (text_to_video), image_to_video (image_to_video) (backend-dependent) - generate_video_batch, image_to_video_batch (repeatable orchestration with explicit seed planning) - generate_angles (multi_view_image) (API exists; no built-in backend implements it yet)

Built-in backends implement: - Images: Diffusers, stable-diffusion.cpp, MLX-Gen, OpenAI-compatible HTTP (../src/abstractvision/backends/) - Current local policy: MLX-Gen supports curated q4/q8 image presets, validated Qwen structured control and Qwen/FIBO masked edits, official FIBO image snapshots, SeedVR2 upscaling, shared LoRA adapters, and Wan 2.2 TI2V/A14B video. This release is validated on Apple Silicon first; the MLX-Gen install extra also exposes Linux support when upstream mlx-gen / mlx markers are available. Local Diffusers text_to_video is experimental and temporarily disabled from normal local surfaces. - Video: - MLX-Gen for Wan 2.2 local text_to_video and first-frame image_to_video - OpenAI-compatible HTTP for optional text_to_video / image_to_video when endpoints are configured (openai_compatible.py)

If you’re looking for “what can model X do?” or “which adapter defaults/compatibilities are packaged?”, the packaged registries are the source of truth: ../src/abstractvision/assets/vision_model_capabilities.json (loaded by VisionModelCapabilitiesRegistry in ../src/abstractvision/model_capabilities.py) and ../src/abstractvision/assets/vision_adapter_capabilities.json (loaded by VisionAdapterCapabilitiesRegistry in ../src/abstractvision/adapter_capabilities.py). Exact runtime route truth still comes from the selected backend. The curation and cross-platform support policy for those registries is governed by ADR 0005.

Internal engineering notes

docs/adr/ contains durable engineering policy.

docs/backlog/ is an internal log (planned work + completion reports). It is not the normative user documentation surface.

Project