Qwen Image is literally unchallenged at understanding complex prompts and writing amazing text on generated images. This model feels almost as if it's illegal to be open source and free. It is my new tool for generating thumbnail images. Even with low-effort prompting, the results are excellent.
Added 2025-08-10 15:50:15 +0000 UTCThis tutorial literally shows how these images were generated with Gemini 2.5 Pro made prompts :
Gemini 2.5 Pro is freely available on Google Studio AI
All images generated in easy to use SwarmUI and they are unmodified raw generations
SwarmUI and ComfyUI install tutorial :
Comments
3:14 Tutorial: How to Automatically Generate Prompts from Reference Images : https://youtu.be/R6h02YY6gUs
Furkan Gözükara
2025-09-04 21:58:36 +0000 UTCyou insinuate that you're going to show how to generate prompts with Gemini 2.5... WHERE!? "This tutorial literally shows how these images were generated with Gemini 2.5 Pro made prompts" - literally, where?
Taiga
2025-09-04 11:56:34 +0000 UTCI had to remove sage attention on the ComfyUI backend in SwarmUI, otherwise Qwen only generated black images for me.
Hockey
2025-08-22 17:37:20 +0000 UTCyou are probably right. it looks like not very much impacted by seed. i still get different outputs but not very wildly different
Furkan Gözükara
2025-08-12 21:13:44 +0000 UTCQwen sticks very close to the prompts. The downside to this is that - in my experience - different seeds do not really change the images as before. So multiple generations almost look the same. You need to change the prompt entirely. Or is there something I'm missing out?
Philipp Capetian
2025-08-11 14:59:31 +0000 UTCjust wait until you see my loras on Qwen :D
Furkan Gözükara
2025-08-10 18:58:53 +0000 UTCNooo, it's not.... The king of image models was already here, hiding! Its good but it's nowhere near as good as Wan 2.2 for images - Not even close - You're missing out! Not many people know it for images, but I have done rigorous testing and it blows Qwen out the water - Well actually I I will caveat that in saying ... It all depends on what you are looking for , Qwen is better for fantasy style images, It's not good at photo realism, it makes everything too polished and "fantasy" like... Which all the examples you made are. BUT for Photo realism, WAN is the best model of all time. PLUS and its a HUGE plus, Wan is soooooo good for making Loras, it comes up with perfect likeness even with small bad datasets. AND the Loras can be used for making brilliant quality videos. So And using the BEST Lora trainer I just found, AI-Toolkit, which is so easy to use and has an amazing GUI, making them is jut great fun. Anyone looking for Photo real fully uncensored images with the best Loras, check out Wan for images is really the king.
Krush Oldwin
2025-08-10 18:17:17 +0000 UTCthey said they will publish such model hopefully. yes training is my #1 priority
Furkan Gözükara
2025-08-10 16:49:05 +0000 UTCCan't wait to see a fine tuning tutorial for this one!!! Also, do you think it can change objects as kontext and redux?
Daniel Cardona Ramirez
2025-08-10 16:11:45 +0000 UTC
