Test comparison. Minimax h3 vs Qwen-Image-2.1. Image generation.
I’d like to thank you for the model.
Despite the feedback, I’m glad to see new models being released openly, and I wish you continued development and hope they keep bringing us great new models.
However, I decided to compare Qwen-Image-2.1 with Minimax H3 in terms of editing.
I’m not sure how reasonable it is to compare these two models, considering that they use different technologies, but since both can be used for image editing, I thought it would be interesting to compare them specifically in terms of character generation.
Both models were run in 8-bit with ConvRot.
GPU: RTX 5090.
For Minimax H3, I used the Minimax H3 Image nodes, which allow image generation, along with Minimax VDN Turbo at 8 steps.
Beta, Euler, 8 steps.
For Qwen, I used a clean workflow.
CFG: 1.0, steps: 25.
Resolution: 2K.
I generated several different characters and used them as references.
The result.
*Image 1: the image source (on the right);
*Image 2: the image with the character (on the left).
It seemed like MiniMax was winning. I don't know what the reason is: the text encoder, the number of parameters, the prompts, or the architecture as a whole.
At least in my experience, when trying to create a consistent character, Qwen-Image-2.1 seems to simply paste the face or the edited parts onto the image, which makes the result look quite crude.
Quite often, the inserted part doesn’t match the original image in terms of either scale or color.
No matter how much I tried changing the prompt, it was difficult to put the character into a different pose or change the camera angle.
On top of that, occasionally there are still various kinds of issues with limbs.
I can't deny that maybe it needs a special approach in terms of promting, and maybe I don't understand it, but I've tried both small and long prompts.
The result is approximately similar, It works better with long prompts, but it still tends to gravitate too much toward the original image.
I also noticed that it doesn't work well with character references, where we give several images of him from different angles at the entrance. Starts duplicating character.
But, If you use only one image, it tends to gravitate toward its angle, lighting, and pose.
Perhaps it will perform much better for other types of tasks.
On the positive side, I found that it works quite well as an upscaler and enhances image detail.
Nevertheless, thank you for releasing the model weights openly.
Image generation models like GPT Image, Nano Banana, Grok Imagine, and others usually require pretty long and detailed prompts. Because of this, the system often routes the prompt through a backend LLM to refine and expand it before feeding it into the image generator. Video models, on the other hand, are trained to take much shorter prompts, but the model itself is still capable of filling in the gaps and generating a complete result. (How should I best explain this concept?)
can you provide your image wf for minimax? I'd like to test it
Yes, if Qwen image 2.1 released in April or May, it would be a good model. but in September, it already far behind.
2.1 can handle large resolution pictures, that's an improve from 2.0, but others are not so good.
Though minimax is almost 4 x the params
Felt the same. Qwen image 2.1 seems to lack this knowledge of physical coherence. Minimax H3 has this physical casual chain built in. Thats why qwen image 2.1 requires dissected prompts touching every bit of influence to work relatively reasonably well.
can you provide your image wf for minimax? I'd like to test it
I can send you the workflow for minimax that I used to generate it.
It’s quite simple.
You can experiment with it yourself.
Here’s the image with the metadata, open it in Comfy.
Dear community,
may I ask for some help, please! I downloaded the above workflow in order to give it a try but I cannot install "taeh3_decoder.safetensors" (I always get a red boundery". I put the file "taeh3_decoder.safetensors" into "ComfyUI/models/vae_approx/" I am pretty sure that I am doing a big mistake but since I am a newby in the world of ComfyUi I don´t no what to to now. Any help is greatly appeciated! Thank you! Edit: If this of concern: I am using ComfyUI portable 0.37