workflow?
Please add workflow.
https://huggingface.co/qpqpqpqpqpqp/Ovis_Image_7B_fp8/raw/main/image_ovis_image.json
Thanks bro, and thanks for the fp8 model !!!
This model sorta drowned a bit in the Z-Image "noise", but its pretty good, and for text renders it does a great job
Official ComfyUI workflows : https://docs.comfy.org/tutorials/image/ovis/ovis-image
(should add at default templates inside Comfy if you ask me ;-))
This model sorta drowned a bit in the Z-Image "noise", but its pretty good, and for text renders it does a great job
Official ComfyUI workflows : https://docs.comfy.org/tutorials/image/ovis/ovis-image(should add at default templates inside Comfy if you ask me ;-))
I'm curious, is it better than Z-Image? Because it appears to me that both are excellent with text
I'm curious, is it better than Z-Image? Because it appears to me that both are excellent with text
Its quite similar in some ways (and borderline from the same company : Alibaba,but different labs, if I remember correctly).
And similar in size.
What made Z-image go "viral" was perhaps releasing the turbo model first, so Z-image is extremely fast (in comparison to Ovis)
Both do good results, my own personal opinion/experience is perhaps that Z-Image is better at usually making good results (a bit Flux ish, always nice poses, faces etc, more forgiving to what you prompt it will do something nice regardless ), while Ovis is a bit more "you gotta prompt well, tweak it" to get the best results. For text I would perhaps give it to Ovis (but in fairness, not done any extensive testing and comparisons)
For text I would perhaps give it to Ovis (but in fairness, not done any extensive testing and comparisons)
I tested both extensively for text, and Z-Image wins. Z-image is smaller, works in only 9 steps compared to 50 steps for Ovis, but at just 9 steps: Z-image has MUCH more accurate spelling, better font-description following, better kerning and spacing, more resolution flexibility, and more consistent performance. Z-image wins regardless of prompting strategy, resolution, sampler, CFG, or shift value. I couldn't find a single point to give to Ovis no matter how hard I tried. :-( Ovis will regularly generate nonsense glyphs instead of real letters, and will change fonts and letter-sizes / line-heights throughout the text passage. It's just really bad at text. (Flux2 is currently the best open-source text model.)
