Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
Apple Silicon video choices
Video Gen offers these complementary paths:
| Choice | Purpose | Important distinction |
|---|---|---|
| FastMetal 1.3B | Small previews | Wan 2.1 sampler |
| FastMetal 5B | Balanced drafts | Dedicated Wan 2.2 sampler, trained three-step schedule |
| FastMetal 14B | Larger FastMetal model | Wan 2.1 sampler |
| FastH3 Preview INT4/INT6/INT8 | Four-step joint video and audio | Dense attention; official source is converted locally |
| FastH3 V2 INT6/INT8 | Eight-step joint video and audio | 80% VSA with reference attention; text-to-video only. Image-to-video stays on MiniMax H3 MLX |
| MiniMax H3 MLX 4/6/8-bit | Original H3 reference | Dense attention, very slow; smaller quantization is not a speed guarantee |
The V2 choices share the official source snapshot at
3da2ddfe1954d9cda4c05b643dc0f26007a655c5, approximately 147.9 GB total.
INT6/INT8 describe the locally converted transformer, not the total download.
Conversion retains routing weights with --include-vsa and uses a separate
cache from dense conversions. An older runtime fails with an update instruction
before conversion. FastVideo is pinned to commit
430e52154e76b902c3cc17a16b3edc1fad790012; older checkouts appear as needing
upgrade in Media Gen Settings. Repair / Upgrade checks out that revision and
prepares the small FastMetal preview decoders, without downloading the large
model snapshots or calling an AI provider. Update FastVideo there before using V2.
A process-local adapter reads video and audio scheduler shifts from the
checkpoint before inference, including V2’s video shift of 10; the denoising
ladder and converted AdaLN cache use the same shifts and eight-step schedule.
The adapter corrects the upstream converter’s four-step cache default before
it strips the original projections. Conversion caches include the schedule
version, step count, scheduler configs and inference-sidecar digest.
The adapter reads the exact trained DMD rungs from fastvideo_inference.json
for both conversion and inference rather than using MLX’s uniform ladder.
Missing, mismatched or unsupported contracts fail before rendering.
SIMD attention remains opt-in upstream; these choices use the reference path.
V2's own weights do accept a first or last frame in ComfyUI (MiniMaxH3ImageToVideo on FastVideo/FastVideo-FastH3-Comfy). The pinned MLX entry does not, and no other MLX checkpoint is this student. Image and first/last-frame requests stay on the MiniMax H3 MLX rows. The search and the contract to wait for are in FastH3 V2 image-to-video.
The Preview INT4 community repack is a preconverted package from MrMofer;
the official Preview INT4 choice converts the source checkpoint on first use.
Their repository links identify the provenance. Existing model IDs and custom
registry choices are preserved. New choices use the registry's existing
_shippedDefaults append mechanism; migration 398 also handles older registries
that predate that marker. Deliberately removed choices stay removed.
The original H3 quantizations share the same conditioner and VAE downloads. They retain the conservative 128 GB memory requirement until end-to-end measurements justify a lower floor. Transformer file size is not peak RAM.
For MiniMax H3 image-to-video, start with a motion-only prompt and a model-native resolution, then inspect the whole clip. A prompt describing a different environment can replace the source scene: image conditioning does not guarantee continuity. In a controlled synthetic comparison, motion-only clips retained the scene through five seconds, while a conflicting snowy-forest prompt replaced it. These observations cover one source and seed, not a guarantee for arbitrary images. See the sampled frames and validation details.
Animation presets now include Rotoscope cinematic and Unreal stylized cinema. These describe a visual treatment in the generation prompt; they do not run an Unreal renderer or perform literal tracing of source footage.
Sources
- FastVideo upstream
- Official FastH3 V2 checkpoint
- Community V2 MLX INT8 example
- FastMetal 5B
- PipeNetwork H3 MLX runtime and quantizations
Upstream timings vary with resolution, frame count, decoder and hardware. PortOS does not treat a published benchmark as a local ETA or promise real-time generation.
Preview to final in the UI
The model picker’s Choosing a model: preview → final help distinguishes small draft renders, a new generation on a quality model, declared Finish pairs, and the gallery’s Upscale action. Cross-model seeds do not promise the same composition. LTX-2.5 generative upscaling can change details and needs its model pack; resizing preserves content without generating detail.
General presets include 384×256 and 256×384 small previews. FastH3 retains its 832×480 quality-check preset and offers a 512×288 test-size button in the help. The latter is experimental, not a validated quality preset. Use the shortest available duration and keep the trained four/eight-step schedule. Smaller canvases reduce denoising work, not download, conversion, or model-load costs.