@helenfromua

Joined June 2026

Posts · 10

Qwen Image 2.1 reference images fail with Float/BFloat16 dtype mismatch

Hi — Qwen Image 2.1 generates normally from text alone, but every generation fails as soon as I add an image reference. Error: RuntimeError: expected scalar type Float but found BFloat16 The traceback consistently goes through the Qwen3-VL vision encoder: transformers/models/qwen3_vl/modeling_qwen3_vl.py get_image_features → self.visual(pixel_values, grid_thw=...) ... torch.nn.functional.layer_norm RuntimeError: expected scalar type Float but found BFloat16 Environment: Windows Pinokio / Maestro NVIDIA GeForce RTX 5090 Laptop GPU, 24 GB VRAM Qwen Image 2.1 7B Loaded text encoder: Qwen3-VL-8B-I

MaestroBlizaine/Maestro

#bug #qwen-image-2.1

@helenfromua

1mo ago

Director Mode... for directors

In my recent time messing around with director mode, I noticed that it really doesn't like preserving the references that I give it. More specifically, it prioritizes taking a generic suggestion from me and making the characters wear different outfits, have slightly different looks, different environments, even if I say that the environment should be kept from specific pictures, or that the characters should be wearing the same outfits that they're wearing in their respective reference pictures. This is why I would like to suggest a director mode skill that is extra faithful to what I tell it,

MaestroBlizaine/Maestro

#director-mode #discussion

@helenfromua

1mo ago

It just broke...

What happened? I was excited to finally be able to extend videos using MiniMax H3. Choose from the dropdown, switch the model from the 4-step one to the full one, and started getting this no matter how many times I restart, hard refresh, try to update. I cannot afford a reinstall because I have a project in motion. Luckily, I can access it through WanGP, but they are linked, and I cannot afford to lose anything. Please figure this out. This literally happened because of one dropdown selection. Steps to reproduce Update to the latest version Switch the model from H3 4-step to H3 Omni Pruned in

MaestroBlizaine/Maestro

#bug #crash #ui-bug #ui-crash

@helenfromua

2mo ago

MMGP preventing startup or slowing down drastically

What happened? a module named MMGP is preventing startup, an upgrade doesn't fix it Steps to reproduce Start Maestro through Pinokio 50% chance of getting a fatal error Your system (OS / GPU / RAM / VRAM / etc.) Windows 11 Home RTX 5090 (laptop) 64 GB RAM 24 GB VRAM Logs / full error output [Maestro] Installing download stall protection... [safe_download] tqdm download hook installed [Maestro] Importing WanGP engine... ██ Detached from Shell 817173c7-d634-4569-94a3-3401c87a950e Incorrect version of mmgp (3.7.6), version 3.7.12 is needed. Please upgrade with the command 'pip install -r requirem

MaestroBlizaine/Maestro

#bug #fatal-error #startup #wangp

@helenfromua

2mo ago

Video extension with Minimax either doesn't work or isn't implemented properly?

What are you trying to do? Extend a MiniMax (FLFV) video with FLFV accordingly in the Extend tab in Maestro. It generates an entirely new video instead with no preservation. Omni doesn't have extension capability at all, it seems. Leads me to believe the Extend tab is either only a "ghost" or I'm still doing something wrong. Cuz how can sliding windows be a thing but not video extension? What have you tried? Adding extra paragraphhs saying things such as "Preserve the source video exactly and generate a continuation smoothly from its end." Setting the duration to be both of the added fragment

MaestroBlizaine/Maestro

#extend-video #minimax-h3 #question #sliding-windows #video-extension

@helenfromua

2mo ago

Minimax H3 video jumps to maximum duration

What happened? Randomly after setting a custom duration, after some time, the slider jumps to the maximum 14.4 point without any influence from me. Steps to reproduce Use Minimax H3 Omni Set the scene Try to set a video duration. Click around, enhance the prompt, whatever else, wait... At some point, the duration sets itself back to the 14.4s maximum. Your system (OS / GPU / RAM / VRAM / etc.) Windows 11 RTX 50590 (laptop) 64GB RAM 24GB VRAM Maestro 1.8.1 Logs / full error output none, not an error. just some weird glitch. Paste the full log output here

MaestroBlizaine/Maestro

#bug #frontend #glitch #minimax-h3 #ui #webapp

@helenfromua

2mo ago

Suggestion: add some GGUF finetunes of Minimax H3

The title speaks for itself. It would be a great addition, in my opinoin, to boost speed and reduce resource usage on consumer hardware. Some variants such as the Q4, Q3 ans Q2.

MaestroBlizaine/Maestro

#discussion #feature-suggestion #idea

@helenfromua

2mo ago

WanGP backend version compatibility and plugins

Hello, Maestro community! I just noticed that this app comes with WanGP v10 as the backend. I was wondering whether it would be possible to upgrade the backend to the latest version instead and whether that could break anything in Maestro. I really like that this app is basically a two-in-one, but I'm not sure why it isn't running on the latest WanGP engine. I also wanted to ask what would happen if I installed a plugin like LTX Director into it, or a model like Ideogram 4. Thanks in advance! LenaR

MaestroBlizaine/Maestro

#backend #customization #question #wangp

@helenfromua

3mo ago

ID LoRa support and DramaBox bug fixes

Hi, Maestro community! I recently switched from WanGP to Maestro for a cleaner, no-nonsense interface that could help me sty focused on creating rather than adjusting a bunch of sliders. As I understand though, Maestro is still very fresh, and still lacks some of the features of its core inspiration WanGP. One of them is the lack of some IC LoRas, such as the ID LoRa for voice reference, which I used to use a lot! Another thing I've noticed is that whenever I switch tabs (video to audio especially), if I don't clear my video inputs from the last generation, the audio models, such as DramaBox,

MaestroBlizaine/Maestro

#feature-request #suggestion #tip

@helenfromua

3mo ago

Local story plot generation

Hi everyone! I'm a small creator who uses AI to make characters for stories and music videos for the songs I write. Today I discovered Maestro and I've officially switched from WanGP now. I'm used to working in the cloud, so its sleek interface is just what I needed to feel at home. I also noticed the Director tab, which is supposed to let me make AI stories and MVs start to finish. While it looks intriguing, it seems to be advertized as API-based, yet in the new UI, I don't see any place to put in API keys for the feature. I wanted to clarify whether this can also use a local LLM, and if that

MaestroBlizaine/Maestro

#question

You’re all caught up.