@warpedtrekker

Music Video workflow - need independent shot/scene timing within fixed audio duration

I'm trying to understand how the new v2.0 Studio workflow is intended to handle music-video editing compared with the previous manual workflow. For an example, I'll use 30-seconds music audio track. My source audio might be 30 seconds, but creatively I may want 4–6 completely different scenes during those 30 seconds. For example: 0:00–0:06 — Woman singing on apartment couch 0:06–0:11 — Woman walking through hallway 0:11–0:20 — Woman singing beside the pool 0:20–0:30 — Woman walking outside in the rain Those scene lengths should be determined by the song/lyrics/beats, not by LTX's maximum gener

#discussion

H3 Director crashes when video_prompt is returned as a dict

What happened? When running MiniMax H3 Director with a remote LM Studio LLM (Qwen 3.6 27B), planning completes successfully with 10 shots, but the pipeline crashes before video generation begins. It appears a prompt field is being returned as a dictionary/object instead of a string, and Maestro attempts to slice it as text. "Director Failed: unhashable type: 'slice'" Error snippet: [MusicVideoPlanner] Leet trigger strip skipped: expected string or bytes-like object, got 'dict' [Director] Plan complete: 10 shots, 98.4s total [Pipeline] Planning error: unhashable type: 'slice' orchestrator.py, l

Director mode - Adjust the cut speed option, missing/not implemented?

What are you trying to do? It says to adjust the cut speed if needed. How do I do that in Director mode? There doesn't seem to be a way to adjust it, nor have it use more reference images. 5 clips is not enough at all even for this short experiment. With 3-4min songs, it never has enough decides on enough clips. This software, I'm having to abandon Auto mode right now. Manual mode at least as more options, but it doesn't have an option to have multiple window segment times. Only fixed seconds throughout the song, which makes cutting shots inaccurate. I have to go back, export part of the audio

LTX 2.5 repair/generation fails in SageAttention because input tensors reach sageattn2 as float32 instead of FP16/BF16.

What happened? Tried generating in Director Mode. Used a compatible image LoRA and reference images for character and environment. Using Flux2 9b and LTX 2.5. Generation fails. Repair will not work. Still says generation failed. File "shared\attention.py", line 528, in pay_attention x = sageattn2_wrapper(...) File "shared\sage2_core.py", line 744, in sageattn_qk_int8_pv_fp8_cuda assert dtype in [torch.float16, torch.bfloat16] AssertionError: Input tensors must be in dtype of torch.float16 or torch.bfloat16

Has anyone got LM Studio to work with Maestro?

What are you trying to do? In director mode, I loaded Qwen3.6 27b and also tried Qwen3.8 27b. Maestro sends my starting description to the LLM. It processes it. Once its done, I eject it because it already sent back the prompts to Maestro. But they are all generic image prompts "REFRAME: medium shot | MOOD: steady.". It also doesnt use any of my additional reference clips. Only that very first main reference clip. So it puts the character in that same scene, identically. I'm not sure what the issue is, if LM studio and the Qwen model is not sending each starting images prompt back, but looks l

Director Clip Regeneration Fails with Float/BFloat16 Dtype Mismatch

What happened? When regenerating individual Director clips with LTX, the rerun reaches Encoding Prompt and then crashes in the Gemma embeddings connector with: RuntimeError: mat1 and mat2 must have the same dtype, but got Float and BFloat16. The model, image, audio/vocal stems, and pipeline all load successfully before the failure. I’ve encountered the same Float/BF16 error previously when regenerating with LTX 2.3, and now with LTX 2.5, so it appears to be a recurring Director rerun/regeneration dtype handling issue , rather than something specific to LTX 2.5. Steps to reproduce 1. Use Direct

LTX 2.3 and 2.5 washed out faded white borders on clips

What are you trying to do? Trying to use ltx 2.3 and 2.5. Either in director or manual mode, some to most clips rendered have faded washed out borders and also a vertical line on very right edge of each clip. Is this a VAE decoder issue or something? What have you tried? Using distilled 2.3 and 2.5. Tried importing my own images or using Flux. Your system (OS / GPU / RAM / VRAM / etc.) 5090

LTX 2.5 + OmniNFT Sync Integration?

Would it be possible in a future Maestro version to offer an LTX 2.5 + OmniNFT Sync integrated checkpoint, similar to the LTX 2.3 + OmniNFT Sync model? The integrated 2.3 version produces significantly better singing lip-sync for me than LTX 2.5 with the separate OmniNFT LoRA. Would OmniNFT need to be retrained specifically for 2.5, or could it be properly merged/adapted to the 2.5 model?

LTX 2.5 terrible lipsync compared to LTX 2.3 +OMNInft sync

I tried LTX 2.5 a lot here on same scenes I rendered in LTX 2.3 +omninft sync. The LTX 2.5 with the added omninft-ltx2.3-22b-rl-lora-r32 does not have the same effect. Why? Horrible lipsync. Sometimes no singing at all. Tried adjusting the values for 2.5 where i had them for 2.3 and same issues.

LTX 2.3 / 2.5 faded white-washed borders in 3-stage render

What happened? Using 3-stage, it appears LTX has an issue with washing out colors along the borders of the video generation. Especially the left and right side. Although i've seen it occasionally with 2-stage as well. But now with 2.5 it is happening a lot. Steps to reproduce 1. Use LTX 2.3 or 2.5 2. Go to multi-shot mode in Studio tab and then advanced and choose 1080p. 3. add still image 4. add text prompt 5. add music clip. 6. generate 16:9 clip. Your system (OS / GPU / RAM / VRAM / etc.) Win 11 / Nvidia 5090 / 64GB / 32GB VRAM / nvme's Logs / full error output Paste the full log output her

LTX 2.5 support?

When can we expect LTX 2.5 to be added? :) I'd like to try the Blackwell distilled NVFP4 version on my 5090. I know it just arrived, but it looks exciting and an improvement over 2.3.

LTX 2.3 color desaturation after first few frames of every clip

Anyone else noticing that when using start images/frames , once LTX is finished rendering the clip, the first few frames quickly lose saturation and stays that way thru the clip end. It's not a Maestro problem since others on LTX forum are seeing it as well.

Multi-Shot LoRA Mode - What IC-LoRA to use?

For Multi-Shot LoRA Mode, which exact IC-LoRA/model does Maestro expect, where should it be installed, and what storyboard prompt syntax should be used for multiple camera cuts within a 20-second generation? Is this feature currently functional with LTX 2.3?

FLUX.2-dev AI Toolkit LoRA Fails to Load in Maestro

What happened? I trained a character LoRA in AI Toolkit using the official FLUX.2-dev BF16 model from Hugging Face. Training completed successfully and produced valid .safetensors checkpoints. When loading the LoRA in Maestro with FLUX.2-dev, the UI shows "Generation failed: Failed to fetch." The terminal reports a tensor split error (split_with_sizes) while loading the LoRA. This appears to be a compatibility issue between AI Toolkit FLUX.2-dev LoRAs and Maestro's current FLUX.2 loader rather than a training failure. Steps to reproduce 1. Create FLUX.2 LoRA in Ostris AI-TOOLKIT using black-fo

Bug Report: Flux Model Loading Failure in Director Automatic Mode

I'm encountering a reproducible error in Director Automatic Mode where image generation fails before any clips are created. Affected Model: Affected Model: PI-FLUX.2 Dev NVFP4 32B and maybe other NVFP4 and others. Error: Error(s) in loading state_dict for Flux: size mismatch for double_blocks.0.img_mlp.0.weight: checkpoint: [36864, 3072] expected: [36864, 6144] The Director pipeline then terminates with: RuntimeError: Generation failed Expected Behavior Director Automatic Mode should load the selected Flux model and begin image generation successfully. Actual Behavior The Flux model fails duri

Maestro Studio mode creating duplicate videos for each clip

What happened? When trying to give Maestro only one clip to make in Studio multi-shot mode, it creates two exact duplicate clips of the same thing. I always have to delete the duplicate every time. Steps to reproduce 1. Select Studio mode. 2. Select Video 3. Select Multi-shot. 4. Upload one video clip only . 5. Generate video and soundtrack based on text prompt. 6. Select duration you want. 7. Start generate. 8. You will see two duplicate clips created after you refresh the page to see the clips.

Regenerate only one clip in studio mode?

What are you trying to do? When I am in studio mode and press the small regenerate button at the bottom of a clip, isn't that supposed to just regenerate that one clip? Instead it starts regenerating everything.

PI-FLUX.2 Dev NVFP4 32B fails to load with state_dict size mismatch (FLUX.2 Dev 32B works)

What happened? The standard FLUX.2 Dev 32B image model loads and generates images correctly (although it uses too much system RAM on my machine). When I switch the image model to PI-FLUX.2 Dev NVFP4 32B , the image prompts are generated successfully, but the image model never loads. Image generation does not start, and the terminal reports repeated PyTorch state_dict size mismatch errors. Switching back to FLUX.2 Dev 32B immediately works again using the exact same project and settings. This appears to be specific to PI-FLUX.2 Dev NVFP4 32B , not the prompts or project itself. Steps to reprodu

How does the main reference photo interact with additional character references?

What are you trying to do? Trying to understand if the main reference photo should be the main singer. A full body view of the character front facing? Then in the additional character ref's, is that were you add closeup views of the face, etc? Then for the additional character ref's description, do we just put a tag like "singer" in it? the speaker_00 should also have the same tag "singer" and role of "singing"?

Multi-Shot LoRA Mode

What are you trying to do? Is the toggle just a placeholder for the future? I don't see any way to add a IC-LoRA in Maestro.