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(prompts): refresh stale Qwen-Edit-2509 + Kontext refs in agent system prompt

mrbizarro3978d57d1c39
diff --git a/agent/prompts.py b/agent/prompts.py
index 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, 25
steps, 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 30
steps — 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, or
a trained LoRA file already?"
-## Path A — Qwen-Image-Edit-2509 references
+## Path A — Qwen-Image-Edit-2511 references
The 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. No
training 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 strictly
matter but the prompt should mention each by descriptor matching
the ref order roughly ("a young woman, a man in his fifties, in a
sunlit kitchen") so the model knows which ref is which.
@@ -554,9 +554,13 @@ Workflow:
`ref_image_path`, OR `mode: "keyframe"` with multiple stills (each
already 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 colored
rectangles, 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 our
client.
**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 the
mflux model dropdown, save, then I'll regenerate."
2. Surface the install hint if `family_status.qwen_edit` is false in
the panel — they need to click "Install Qwen-Image-Edit (multi-ref
diff --git a/mlx_ltx_panel.py b/mlx_ltx_panel.py
index 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">