Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.
Python ai-filmmakingai-videoai-video-generatorapple-siliconhailuoimage-to-videolip-synclocal-ailora-trainingltx-2ltx-2-5ltx-videomacosminimax-h3mlxmusic-generationoffline-aipinokiotext-to-videovideo-generation
Phosphene dependency pins: full rationale
Moved out of CLAUDE.md §4. The long-form rows for mlx and ltx-2-mlx, with all measurements.
| Pin | Why |
|---|---|
mlx==0.31.1 |
mlx 0.31.2 introduced a numerical regression that attenuates LTX 2.3 vocoder output by 0.95 * hw.memsize, identical formula on both), so it scales down with the Mac but never stops filling it. /usr/bin/time -l peak, same prompt/seed/dims, MLX memory+cache limits set to each Mac's derived value: Q4 768×432 — 16 GB 17.41 → 20.35 (0.32.2: 19.35), 32 GB 25.66 → 33.18 (97% of physical), 64 GB 25.53 → 54.28; Q8 1024×576 — 48 GB 39.49 → 49.95 (97% of physical). On the 64 GB box the Q4 render moved system pressure 58% → 84% with +1.5 GB swap, the Q8 one hit 87% — past the 82% at which plan_memory_policy() forces the streamed decode (MLX_RESIDENCY_SET_MAX_PCT=0 reproduces the footprint to within 300 KB over three runs. The growth is allocator cache, so an explicit mx.set_cache_limit() is the prerequisite — shipped, on this pin, ahead of any version move: mlx_warm_helper.py now caps the cache at one eighth of physical RAM (floor 2 GiB, ceiling 8 GiB), applied at helper start and re-asserted before every render because the pipelines drop it to 0 on paths we do not take. On 0.31.1 that is Q4 768x432 121f 25.53 -> 16.50 GB for 71.45 -> 68.84 s and Q8 1024x576 at the 48 GiB tier 39.49 -> 25.39 GB for 139.84 -> 138.71 s, both sha256-identical to the uncapped render. And the pin cannot move alone — see scripts/check_post_update.js. 2026-08-28, second pass: the cap shipped and the move was REFUSED AGAIN. With the policy applied to both versions, 0.32.1's peak as a share of physical RAM is 16 GiB 101% → 116%, 32 GiB 75% → 68%, 48 GiB Q8 77% → 82%, 64 GiB 37% → 41% — three of four tiers worse than 0.31.1 uses today, and the Comfortable 48 GiB tier lands ON the 82% line plan_memory_policy() treats as pressured. The cap cannot reach it because the excess is not cache: at an identical cap, 0.32.1's active memory is +12.0 GB on Q4 and +29.2 GB on Q8. Cache 0 does get it under (Q8@48 GiB 32.25 GB, Q4@16 GiB 17.04 GB) and erases the entire speed win at the same time. There is no cache setting where 0.32.1 beats 0.31.1+policy on both axes, so mflux stays at 0.18.0 too. |
ltx-2-mlx PINNED to the FORK TAG v0.14.19+ltx25.7 (mrbizarro/ltx-2-mlx = v0.14.19 1192051 + the LTX-2.5 port + the schedule/sampler/live-preview work + the i2v anchor re-pin on BOTH loops + +ltx25.7: IC-LoRA source_video=, the seam Upscale ×2 renders through — the tag's own history is in scripts/pinokio/ltx_checkout.sh) |
The pin is a TAG and it lives in scripts/pinokio/ltx_checkout.sh, which is the ONE checkout implementation both lanes call (install.js dispatches it; scripts/post_update.sh calls it). It was a bare SHA through v3.8.3 and was reachable only by luck — a force-push would have stranded every install with an un-fetchable pin and a dead Update button. check_ltx_pin.js is the gate. The packages report 0.14.19+ltx25.7; _LTX_EXPECTED_VERSION in mlx_warm_helper.py must equal that string or every render logs VERSION SKEW — move the two together. This pin is what lets LTX-2.5 render through the panel, and on dev/beta 2.5 is the registry DEFAULT (2.3 stays installed, selectable via LTX_MODEL_VERSION=ltx23). v3.8.x did NOT register ltx25 at all and shipped 2.3 as the only generation, because the 2.5 weight packs were not publicly downloadable when it was cut — a fresh install would have booted into a generation it could not fetch. They are published now, and v4.0.0 closes that divergence: 2.5 is the registered default everywhere. If you ever need to curate a release back to a single generation, it is one MODEL_VERSIONS dict plus its three required_files.json rows (watch the kind: "base" trap: any base row is mandatory for an install to read complete, in the panel and in pinokio.js). The pin diverged for the same reason: v3.8.x shipped 871694d / 0.14.19+ltx25.1, one commit behind, because e6be9d6 (the 2.5 sampler + stage-2 schedule, below) had no 2.3 regression evidence behind it at cut time. It does now — a 2.3 sha256-identity pair across the change — so the next release should choose the pin deliberately rather than inherit either. The base tag v0.14.19 landed 2026-08-11 (a catch-up from v0.14.8, itself a 2026-06-01 catch-up from v0.14.0). STAY PINNED — this is NOT "main HEAD", and not upstream's tip either. The v0.14.8 bump had pulled dgrauet's native _pre_denoise_flush (the Metal-watchdog fix for the I2V "mosaic" on memory-pressured Macs, #17), budget-aware VAE decode tiling and first-class frame_rate, which cut us from 7 runtime patches to 1 (codec-only). The v0.14.19 bump keeps that shape and adds: checkpoint-read av_ca_timestep_scale_multiplier (0.14.11 — audio/dialogue weighting changes on every render, toward upstream parity, plus LTXModelConfig.from_checkpoint_dir()), frame_rate forwarded at the A2V/lipdub call sites and no more shortest-stream mux truncation (0.14.15), quantized transformers at any group_size (0.14.16), and GPU-watchdog explanation + DiT-freed-before-VAE-decode + a Gemma cache-limit fix (0.14.19). The +ltx25.2 bump (2026-08-12) is the first one that changes what a 2.5 render computes, and only a 2.5 render: (a) the euler-ancestral sampler is finally REACHED — it had been implemented with 48 tests and zero callers, so every prior 2.5 render silently used the 2.3 Euler step, against upstream's own ANCESTRAL_SAMPLER_SINCE_VERSION = (2, 5) and all three official ComfyUI 2.5 templates; keyframe/flf2v deliberately stays on Euler because the flf2v template pins eta to 0. (b) stage 2's first sigma is now the vendor's 0.85, read from video_ltx2_5_t2v.json node 395, where we had been using 2.3's 0.909375 — a refine pass starting at the wrong noise level. Both are version-keyed off the checkpoint's own model_version, and 2.3 output is proven sha256-identical before and after. No model re-download: 0.14.13+ prefers a versioned transformer-distilled-*.safetensors and falls back to the unversioned name, and update.js trims the -1.1 variants, so file resolution is identical to v0.14.8. Bumping the pin further is a deliberate, tested operation — read the release notes, bump _LTX_EXPECTED_VERSION in mlx_warm_helper.py, smoke-test the matrix on dev, then bump BOTH install.js and update.js in one commit. The +ltx25.4 bump (2026-08-14, v4.0.2) is the i2v anchor. The conditioning item was always built correctly — frame_idx=0, strength=1.0, mask_value=0, a hard latent replace — and both loops forced those tokens back into the model's x0 estimate every step. Under Euler that was sufficient by arithmetic accident: velocity (x - x0)/sigma is exactly 0 at a pinned token, so the step returns it unchanged. The euler-ancestral step 2.5 selects has no such property: it rescales every token by alpha_next/alpha_down and adds fresh Gaussian noise, unmasked. Measured on the real 2.5 schedule the anchor is halved and buried under noise 0.86x its own scale after step 1, so every forward from step 2 on saw garbage where the image should be and the composition was decided without it; the terminal step stamped the image back in, which is why the delivered frame 0 always matched while the clip was a different shot. The fix re-composites the sample against clean_latent/denoise_mask after the ancestral step at all three sites, guarded by the existing uniform-mask flags — the other half of what ComfyUI's 2.5 i2v graph does. Proven, not argued: t2v at a fixed seed is sha256-identical before and after (fbf09533583c423ab9ad3c0a437aee3484f15c0c5cceb2b1fc4ecc6d59071abb), and frame-vs-input normalised cross-correlation at seed 424242 moves 0.0934 -> 0.7878 (frame 36) and 0.0902 -> 0.7187 (frame 72) while frame 0 holds at ~0.975. Every Euler lane — 2.3 i2v, keyframe/flf2v with its deliberate eta=0, a2v, extend, restore, ic_lora, lipdub — never reaches the branch. The +ltx25.5 bump (2026-08-15, v4.0.7) closes that on res_2s too, which is the lane 2.5's HQ tiers run: the owner reported a reference image producing a high-quality clip of something else on High / High-720p. res2s_denoise_loop takes repin_masked_sample (default False = the old bytes) and ti2vid_two_stages_hq resolves it per generation, so 2.5 anchors i2v for real while 2.3 is byte-identical BY CONSTRUCTION — the reason this was deferred in +ltx25.4 is answered by construction rather than by measurement. denoise_loop takes the same flag defaulting True, so passing False deliberately restores the drift: that is Inspire, shipped as an explicit i2v mode (loose_reference through generate_two_stage/generate_and_save on both pipelines, forwarded only when true). tests/test_repin_masked_sample.py pins all four directions. The historic dcd639e (0.1.0) detour broke the Extend cfg_scale API; do not revisit. |