Incredibly good model
This model is the best thing that has come in a long time, I really hope that Lora support comes soon.
i find it amusing again how they try to sell the cumbersome json prompt need as an advantage.
It should better be really good to make me bother with this crap.
@Andyx1976 You can write normally, it is not mandatory to write in .json format. It works great with normal input. I use the node in comfyui fibo2.0 text encoder. You don't have to force photos into it. I think the text encoder converts it directly into this format. I've always written normally, like with any other model. I get incredibly good realistic results; the photos are really great. Just give it a try. Best regards
according to their model page only because of their prompt enhancer (which is of course a censored llm). Not the text encoder (clip), that may expect the json. But should also not be able to change the written prompt.. Is that turned on by default?
It will probably be the same model but used in different ways, usually a text generate node in it.
That's what they say: "Generate Short text prompt The VLM expands your idea into a rich structured prompt, and Fibo 2.0 renders it. You get the image and the full JSON."
(VLM as in LLM but with vision support), not the clip. They don't really say if you can bypass it and actual give it normal text. Ideogramv4 (also json needy) actually responded to that with a in-image refusal.
I got rid of ideogramv4 after a few days, because other models are as precise from a normal text. Just like that. The whole json stuff had no advantage.
So I have tested the model extensively since its release, I have used .json format and compared it with a normal prompt, I get just as good results with normal inputs without .json, I have created my first loras that work perfectly and have perfect likeability. I'm more than satisfied with this model and really excited about it. I'm currently using the dev version with the Turbo Lora and CFG 3-5. The quality is great, and it works great even at higher resolutions. Just forget about this.json format and just test the model.
I have. And i agree on the prompt. I strongly disagree on the resolution:
any edit/generation job i tried fails on resolution higher than 1 megapixel. It does everything correct at 1mpx. At 2 it makes tons of mistakes. Not messy anatomy, so if you don't care about precision, the results are still usable. But don't try anything precise like text or specific amounts.
Maybe a shirt color swap or so works fine at 2+ but anything more ambitious is full of errors.. Also the image quality isn't great at 2mpx, but that maybe my settings (but maybe not). The mistakes are deffo not, that is the model. It is good at implementing the guy from the reference image though.
using default comfy template workflow, with prompt enhancer (same issue without). both turbo, prompt at the end:
1mpx, almost perfect, and quite nice:
2mpx , EVERY run too many people, too many letters, too many combine harvesters... some or all of them, but never correct. All precision of a editing model is gone.
dev+turbo, same thing. perfect at 1, mistakes at 2mpx (this):
4Mpx, here also anatomy fails (hands between left two guys)
To compare: Flux3 (distilled i.e. turbo,) at 8 (!)Megapixels. All seems correct.:
You know, let's live a little.. Flux3 at 16 (sixteen) megapixels, needs more steps even in turbo. But there are no errors i can see.:
prompt (unenhanced):
Three soft fluffy white clouds drift across a bright sky, their rounded edges subtly shaping the letters “S”, “F”, and “W” in gentle negative space, hovering above a vivid meadow filled with colorful wildflowers spread across green grass, in the distance one lonely combine harvester, natural daylight, airy and whimsical mood.
Below walk three young man aged 18, aged 19 and aged 40, all wearing differently colored, simple dress shirts and jeans and heavy black boots.
the left and central man walk chippy arm in arm, while the right man walks slower behind them.