Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More
26 contributions to Muse Collective
Tip for all text encoder victims!
Perhaps some of you have also encountered the issue where Minimax fails to follow your prompts. Regardless of which Minimax model I tried, 70% of the time the output didn't match what I had requested—it was driving me absolutely crazy. Yesterday, I stumbled upon a YouTube video by Mickmumpitz, which revealed that the text encoder I had been using (`qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`) is actually designed exclusively for Nvidia Blackwell-series graphics cards. Since I’m still running an older 3090 24GB card, I decided to switch to a different text encoder (`qwen3vl_32b_minimax_h3_int8_convrot.safetensors`). Since making that change, the results now align with my prompts 90–95% of the time.
1 like • 4d
Hmmm. Well, I have a blackwell, and now according to you, it appears I've been using the wrong text encoder (The one you just switched to...!) Although, I generally go for ones that work for NSFW and just leave them in place during SFW stuff too. Time to dig deeper, thank you.
FREE Video Ref Test Workflow
Hey guys 👋🏻 Ok so I have been playing with Video Reference in the new Director V2.1 and thought I would share the test node I was using. Take a look at the video below 👇 I would really love to see some examples in the community, so take this as a challenge! The Github repo is below and the link to the workflow is attached. Have fun!! GitHub: https://github.com/muse-collective-26/Muse-Video-Swap-H3
FREE Video Ref Test Workflow
1 like • 5d
Me: "I'll leave a batch running overnight with slightly different prompts and different seeds and check the results in the morning." Windows Update: "I don't think so..."
0 likes • 4d
Working well when correctly prompted (amazed at the difference the seed makes too considering how little effect the seed has on some models lately). Love the way the seed will make it perfect, and then change it, and the guy is doing a tiktok dance (Have you triggered some secret level in the models?) Now... Any chance of merging this into a continuation/Extend? :D Otherwise I'll try to merge myself into other extend-workflows I have.
Video Editor Mode!
Ok, seriously, I am going to have to start charging more for this soon!! I am adding video editor mode. It will be a separate mode where you can drop a video and edit it or use it. See below. One clip has the character change sword position, another changes the character and the last changes character and scene.
Video Editor Mode!
1 like • 7d
Nice. Everyone talks about character replacement, am I the only one who wants environment replacement? In my mind that is a little easier as it is mostly static? Or is that an oversimplification? Also, how long at the clips that you can edit? (I have 5090 with 32gb and 128GB ram, guessing it is based on that somehow?)
1 like • 7d
@Andy Whittaker Ah, I see. I thought with a reference image or ref-mod for the background it would just 'work' :D I've tried just about every method for character replacement and found just 1 that actually worked (joyAI i think it was called for LTX, and that only worked well in the Wan2gp environment, and not in comfy, therefore painfully slow. Then again, I don't really need or use character replacement, so I don't spend a lot of time on it. I find it a bit terrifying to be honest along with face swap... That said, would love to bring some video-game characters to life in short scenes)
Upscale video option in the chuck section
Is it possible to have a upscale function in the chunks section so if your happy with generated chunks you can choose to upscale a chunk or all chunks for your final video output. Example, first run generated at Megapixels: 0.6, quicker generations. Upscale option that is driven from the Upscale Model,Upscale Sizing,Target Megapixels setting that are already in the workflow. This will allow quicker generation times with the option to upscale when you decide to.
Upscale video option in the chuck section
1 like • 7d
Genuine question: WOuld you not prefer and external upscale? I've been running a basic LTX upscale, it is fast and does a great job, much better than RTX, and I thought faster than Minimax. I've been using the director to build the basic, than an LTX to upscale and frame insert.
What do you save before updating ComfyUI?
For those running several custom node packs, what has actually let you recover an older working setup after an update broke it? A second install, a full folder backup, something else? Interested in what you had to restore beyond the workflow JSON.
3 likes • 9d
1) I have a basic install of Comfyui that I use ONLY for loras, diffusion models, checkpoints, encoders etc. Basically just a dumping ground that is never updated. 2) then I have my main Comfyui that is the one I work with and have the custom nodes in. 3) I back up this comfyui regularly, because i know I'm only one "install missing node" away from disaster at any time. 4) If if do need to restore or whatever, I can just do an install, and copy the Customnodes directory accross and I'm all set ( along with the extra_model_paths.yaml of course)
3 likes • 8d
@Rob Bertrand my launch bat has this --disable-auto-launch because I only open it in Chrome, which isn't my default browser. Also means I can just go to 127.0.0.1 for a clean workspace with no old/broken workflows.
1-10 of 26
Sjt Jt
3
24 points to level up
@sjt-jt-4213
Not today thanks.

Active 4d ago
Joined Oct 16, 2026