Phosphene 3.7.0 is out - the workflow release for Hailuo H3.
Drafts in about a minute. Draft tiers now decode through a tiny fast decoder instead of H3's full 36-layer decoder - the single biggest fixed cost in a render. A 640x384 5-second draft lands in roughly a minute; a full 10-second single-pass draft measured at 5:05 on an M4 Max. Delivery tiers keep the untouched decoder, so nothing you ship is affected.
A real 10 seconds. "10s single pass" now sits next to the chained 10s on every canvas up to Native 1344x768. One window: no seam, no repeated action after 5 seconds, no drift between halves. This was the top community request after the chained 10s reports - thank you @Wizard_1981 for pushing on it.
LoRAs for Hailuo H3, from CivitAI. Browse and install in the same picker LTX uses. H3 has one adapter slot, so a LoRA and Turbo can't both run - the panel says so and lets you choose instead of silently dropping one. ComfyUI-namespace files are converted losslessly on the way in; diffusers/PEFT files are refused with the reason (their scale factor is not in the file - guessing it renders noise). Validated end to end with a community character LoRA: every one of its 208 modules applied.
Same seed, same prompt - left is the base model, right is a CivitAI character LoRA through the new picker.
H3 now runs on 48 GB Macs. A quantized Q8 engine renders at 27 GiB peak instead of 43, about 5% slower, with faces holding in a reviewed side-by-side. RAM picks the engine automatically: 60 GB+ keeps full bf16 quality, below that the Q8 pack is used - and 48 GB Macs, previously told H3 was impossible, build the pack locally at install in about 5 minutes. On a 64 GB Mac you can force Q8 in Settings to keep ~15 GiB free for the rest of your work while rendering.
Also fixed: a nasty one where the Generate button went silently dead after Load Params or Finish (a fractional clip duration made the form invalid - browsers refuse to submit invalid forms with no error at all), and gallery cards now say "10s clip - rendered 5 m" instead of one unlabeled number.
Heads up: Phosphene now sends anonymous usage counts (version, hardware class, render outcomes - never prompts, never media). It tells you on first run and Settings has the off switch. The full schema is in docs/ANALYTICS.md.
Existing H3 users: run "Repair Hailuo H3" from the Pinokio menu once to pick up the new runner - it carries the LoRA slot, the draft decoder and the Q8 tooling.
Thanks: @Wizard_1981, Morac2, edserra, Kijai, larryvrh, LightX2V, madebyollin, Tr1dae.