Four things people reported, fixed together.
Image mode anchors on High and High-720p again. Give those tiers a reference image and you could get a beautiful clip of something else: the second-order sampler they run re-noised the conditioning every step, so the model composed the shot without ever really seeing your image - only the delivered first frame matched. 4.0.2 fixed this class on the fast lanes and deliberately left the High lane open; it is closed now, resolved per model generation, so LTX-2.3 renders are unchanged byte for byte.
Inspire. The old behaviour turned out to be accidentally useful - a reference that guides subject, style and palette while the shot re-composes itself - so it ships as a choice instead of a defect. Image mode on LTX-2.5 now offers Anchor (animate this image) and Inspire (new shot from it), and the control says plainly that Inspire will not show your picture as frame one.
Ingredients tells the truth on LTX-2.5. Its reference adapter is 2.3-trained and no 2.5 version has been published, so on 2.5 your references were quietly ignored and you paid a full two-stage render for an unrelated clip. It now says so and refuses instead of spending your GPU. LTX-2.3 installs are unaffected. For reference-guided work on 2.5, use Image mode with Inspire.
Hailuo H3 renders vertical. Every H3 canvas shipped landscape, which put social formats out of reach on that engine. Portrait is the same canvas rotated - same pixel budget, same wall clock, same tier estimate - and it is a per-render switch, so every clip you already rendered still replays exactly.
Also: H3's live preview now explains its absence instead of showing an empty stage (the panel side is done; no published H3 runner can publish preview frames yet). And the anonymous usage stats got useful without getting personal - render events now carry the same coarse machine class the boot event always sent, a bucketed wall time, and an error class computed on your own machine, so "which failures happen on which hardware" is answerable at all. No prompts, no media, no filenames, no LoRA names. docs/ANALYTICS.md documents every field; the off switches are unchanged.
Engine pin moves to v0.14.19+ltx25.5. No weights change.
Release notes: github.com/mrbizarro/Phosphene/releases/tag/v4.0.7
AboutPhosphenemrbizarro/phosphene
Four things people reported, fixed together.
Image mode anchors on High and High-720p again. Give those tiers a reference image and you could get a beautiful clip of something else: the second-order sampler they run re-noised the conditioning every step, so the model composed the shot without ever really seeing your image - only the delivered first frame matched. 4.0.2 fixed this class on the fast lanes and deliberately left the High lane open; it is closed now, resolved per model generation, so LTX-2.3 renders are unchanged byte for byte.
Inspire. The old behaviour turned out to be accidentally useful - a reference that guides subject, style and palette while the shot re-composes itself - so it ships as a choice instead of a defect. Image mode on LTX-2.5 now offers Anchor (animate this image) and Inspire (new shot from it), and the control says plainly that Inspire will not show your picture as frame one.
Ingredients tells the truth on LTX-2.5. Its reference adapter is 2.3-trained and no 2.5 version has been published, so on 2.5 your references were quietly ignored and you paid a full two-stage render for an unrelated clip. It now says so and refuses instead of spending your GPU. LTX-2.3 installs are unaffected. For reference-guided work on 2.5, use Image mode with Inspire.
Hailuo H3 renders vertical. Every H3 canvas shipped landscape, which put social formats out of reach on that engine. Portrait is the same canvas rotated - same pixel budget, same wall clock, same tier estimate - and it is a per-render switch, so every clip you already rendered still replays exactly.
Also: H3's live preview now explains its absence instead of showing an empty stage (the panel side is done; no published H3 runner can publish preview frames yet). And the anonymous usage stats got useful without getting personal - render events now carry the same coarse machine class the boot event always sent, a bucketed wall time, and an error class computed on your own machine, so "which failures happen on which hardware" is answerable at all. No prompts, no media, no filenames, no LoRA names. docs/ANALYTICS.md documents every field; the off switches are unchanged.
Engine pin moves to v0.14.19+ltx25.5. No weights change.
Release notes: github.com/mrbizarro/Phosphene/releases/tag/v4.0.7