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
fix(prompts): refresh stale Qwen-Edit-2509 + Kontext refs in agent system prompt
mrbizarro3978d57d1c39
diff --git a/agent/prompts.py b/agent/prompts.pyindex fa93bd9..9134c43 100644--- a/agent/prompts.py+++ b/agent/prompts.py@@ -393,7 +393,7 @@ chose.**Engine for photorealistic stills with character refs:** default to`engine_override="flux2_edit_high"` (FLUX.2 Klein-Base-Edit, Q8, 25steps, guidance 4.0 — non-distilled, ~3-5 min/image, photographic-output). Alternatives: `qwen_edit_high` (Qwen-Image-Edit-2509 Q8 30+output). Alternatives: `qwen_edit_high` (Qwen-Image-Edit-2511 Q8 30steps — strong on multi-character composition) and `kontext_high`(FLUX.1 Kontext Q8 30 steps — different aesthetic). The bare`flux2_edit` and `qwen_edit` presets are FAST but DISTILLED — output@@ -501,11 +501,11 @@ stronger because it conditions LTX itself, not just the keyframe still.When in doubt, ASK. "Do you have photos of this character to share, ora trained LoRA file already?"-## Path A — Qwen-Image-Edit-2509 references+## Path A — Qwen-Image-Edit-2511 referencesThe cleanest fix for "same prompt, different seed = different person"is to **anchor identity at the still stage** by composing the-character into every shot's keyframe via Qwen-Image-Edit-2509. No+character into every shot's keyframe via Qwen-Image-Edit-2511. Notraining step, no LoRA file — just 1-3 reference images per shot.The image engine's `qwen_edit` family wraps `mflux-generate-qwen-edit`@@ -544,7 +544,7 @@ Workflow:```refs=["/abs/path/emma.png", "/abs/path/marcus.png", "/abs/path/kitchen.png"]```- Qwen-Edit-2509 will compose all three. Order doesn't strictly+ Qwen-Edit-2511 will compose all three. Order doesn't strictlymatter but the prompt should mention each by descriptor matchingthe ref order roughly ("a young woman, a man in his fifties, in asunlit kitchen") so the model knows which ref is which.@@ -554,9 +554,13 @@ Workflow:`ref_image_path`, OR `mode: "keyframe"` with multiple stills (eachalready character-locked) for cross-shot continuity.-The image-engine's family must be `qwen_edit` for refs to take effect.-Other families (flux1, flux2, z_image, fibo) silently drop refs and-the candidate dict's `refs_ignored: true` flags the no-op.+The image-engine's family must consume reference images for refs to+take effect. Multi-ref (1-3 images via `--image-paths`) families are+`qwen_edit` and `flux2_edit`. The `kontext` family takes a single+reference via `--image-path` (only the first ref is used; extras are+silently dropped). Other families (flux1, flux2, z_image, z_image_turbo,+fibo, qwen) drop refs entirely and the candidate dict's+`refs_ignored: true` flags the no-op.### When you see `refs_ignored: true` on every candidate — diagnose the engine, NOT the photo@@ -567,16 +571,19 @@ that produce this:- `kind: "mock"` (default for fresh installs) — paints flat coloredrectangles, ignores everything.- `kind: "mflux"` with `mflux_family` set to one of: flux1, flux2,- z_image, z_image_turbo, fibo, qwen, kontext — these are text-to-image- families; they don't take reference images.+ z_image, z_image_turbo, fibo, qwen — these are text-to-image+ families; they don't take reference images. (`kontext` accepts a+ single ref via `--image-path`; if you passed it 2-3 refs, only the+ first is used and the others are silently dropped — that is *not*+ a refs_ignored=true case but worth flagging to the user.)- `kind: "bfl"` — cloud Flux, no multi-reference primitive in ourclient.**Correct response when refs_ignored fires:**1. Tell the user *exactly* this: "Your image-engine isn't set to- Qwen-Image-Edit-2509. Open Settings → Image generation, pick- `Qwen-Image-Edit-2509 (Apache 2.0 · multi-reference)` from the+ Qwen-Image-Edit-2511. Open Settings → Image generation, pick+ `Qwen-Image-Edit-2511 (Apache 2.0 · multi-reference)` from themflux model dropdown, save, then I'll regenerate."2. Surface the install hint if `family_status.qwen_edit` is false inthe panel — they need to click "Install Qwen-Image-Edit (multi-refdiff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.pyindex 299c3cc..4a8b113 100644--- a/mlx_ltx_panel.py+++ b/mlx_ltx_panel.py@@ -6033,7 +6033,7 @@ class Handler(BaseHTTPRequestHandler):# Per-family install status — exposes which mflux-generate-*# binaries are present so the browser can show install hints# next to options the user hasn't installed yet (e.g.- # Qwen-Image-Edit-2509). Probing all families on every config+ # Qwen-Image-Edit-2511). Probing all families on every config# fetch is cheap (a few stat() calls).family_status = {}for fam in agent_image_engine.MFLUX_FAMILY_BIN.keys():@@ -14947,7 +14947,7 @@ HTML = r"""<!doctype html><option value="__custom__">Custom HF id / local path…</option></select><div id="agentQwenInstallHint" class="field-help" style="display:none;margin-top:6px;padding:8px;background:rgba(207,150,34,0.08);border:1px solid rgba(207,150,34,0.4);border-radius:6px;color:var(--text);font-size:12px;line-height:1.4">- <b>Qwen-Image-Edit-2509 needs a one-click install</b> — open Pinokio's launcher menu and click <b>"Install Qwen-Image-Edit (multi-ref keyframes)"</b>. ~30 s for the package; weights (~22-34 GB) download lazily on first use.+ <b>Qwen-Image-Edit-2511 needs a one-click install</b> — open Pinokio's launcher menu and click <b>"Install Qwen-Image-Edit (multi-ref keyframes)"</b>. ~30 s for the package; weights (~22-34 GB) download lazily on first use.</div></div><div class="field" id="agentMfluxCustomField" style="display:none">