Skip to content

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.

fix(comments): un-mislead two load-bearing code comments (doc audit H7/H8)

mrbizarroc139a2fbbb67
diff --git a/image_engine.py b/image_engine.py
index 527775a..a20ed5f 100644
--- a/image_engine.py
+++ b/image_engine.py
@@ -1071,11 +1071,14 @@ def _generate_mflux(prompt: str, n: int, width: int, height: int,
# ones that did rather than failing the whole batch. Order by seed
# so candidate numbering is deterministic.
#
- # mflux >=0.18 changed --output handling: even when the template
- # already substitutes {seed}, mflux ALSO appends _seed_<seed> to the
- # filename. So our `cand_{seed}_mflux.png` template lands on disk as
- # `cand_<seed>_mflux_seed_<seed>.png`. Without this fallback every
- # image-gen run since the upgrade returned an empty candidates list,
+ # IMPORTANT (2026-05-31 review H7 — do NOT delete this as "future compat"):
+ # mflux is pinned ==0.17.5 (<0.18). The PINNED 0.17.5 appends _seed_<seed>
+ # to the filename whenever n>1 seeds are passed (the default agent batch is
+ # n=4) — even when the template already substitutes {seed}. So our
+ # `cand_{seed}_mflux.png` template lands on disk as
+ # `cand_<seed>_mflux_seed_<seed>.png`. This fallback is LOAD-BEARING on the
+ # current pin, not dead >=0.18 compensation. Without it every multi-seed
+ # image-gen run returned an empty candidates list,
# which meant: no sidecar JSON written (the for-loop ran zero times),
# users saw thumbnails in the gallery via the library scan but had
# no per-image metadata to show / reproduce / debug. The candidates
@@ -1087,7 +1090,7 @@ def _generate_mflux(prompt: str, n: int, width: int, height: int,
path = Path(output_template.format(seed=seed))
if not path.is_file():
# Fall back to scanning the output dir for the EXACT filename
- # mflux >=0.18 writes (template + auto-`_seed_<seed>` suffix).
+ # pinned mflux 0.17.5 writes on n>1 (template + auto-`_seed_<seed>`).
# Substring matching by `seed_str in p.name` was wrong: seed
# 123 would also match `cand_9123_mflux.png` from a previous
# run in the same agent shot folder, returning the wrong PNG.
diff --git a/mlx_warm_helper.py b/mlx_warm_helper.py
index 086465b..58d849e 100644
--- a/mlx_warm_helper.py
+++ b/mlx_warm_helper.py
@@ -1228,11 +1228,18 @@ def _generate_latents(pipe, *, needs_image: bool, kwargs: dict):
# Upsampler ≈ ~30s constant
# Total ≈ 5.0 step-equivalents + 30s vs 8.0 step-equivalents native.
#
- # OFF by default for v3.0.x stability — flip on for A/B with
+ # DEFAULT = single-stage, by Mr Bizarro's explicit call (2026-05-27).
+ # The A/B was run (tasks #49/#51, the e1_ab/compound_matrix HTML reports):
+ # two-stage is ~30-35% faster (M4 Max 7:40 → ~5:15) BUT visibly SOFTER —
+ # fine detail is reconstructed during the 2x upsample + 3-step refine
+ # rather than denoised directly at full res. Mr Bizarro eyeballed both and
+ # PREFERS single-stage's sharpness, so single-stage stays the default.
+ # (The 2026-05-31 deep review flagged single-stage as upstream-"OOD"; that
+ # concern is theoretical — the measured A/B showed single-stage sharper,
+ # not artefact-broken, on real prompts. Do NOT flip the default to
+ # two-stage without Mr Bizarro re-deciding.) Two-stage remains available
+ # opt-in for users who want the speed and accept the softness:
# PHOSPHENE_T2V_TWO_STAGE=1
- # Same prompt+seed twice (env on / env off) to compare. The eventual
- # Option-A (deleting this legacy branch entirely so two-stage T2V is
- # universal) lands after we've validated quality on dev for a stretch.
_t2v_two_stage = os.environ.get("PHOSPHENE_T2V_TWO_STAGE", "").strip().lower() in ("1", "true", "yes", "on")
if not needs_image and not _t2v_two_stage and hasattr(pipe, "generate"):
try: