Quick Thoughts on HiDream-I1 & E1
Amazing output but quite a beast to work with
The new HiDream-I1 17 billion parameter model was released just couple of weeks ago, quickly followed by the E1 model and it has already gained significant popularity. HiDream-I1 comes in three flavours - Full, Dev and Fast. The Full model weighs in at over 32GB and is too much for a 16GB VRAM graphics card.
While I have seen others achieve good results with the Dev model, my current go to is a Q8 quantised version of the Full model, which is 18GB. It runs on my 16GB 4060Ti OC card at about 7s per iteration and produces some stunning quality images.
As well as the model file you will need the text encoders and a workflow - this is well documented on the ComfyUI website where you can also find links to the quantised GGUF models. I would also recommend Tenofas’ HiDream workflow.
As the model is still very new there are few LoRAs or refined models yet, but the model is pretty amazing just as it is, provided you are patient as a 50 step run is much slower than Flux!
A major difference compared to Flux is that HiDream allows for four text encoders, each having a different text prompt. This is somewhat confusing but aims to facilitate tuned prompts for different encoders. The argument for this being for example that Clip encoders are better with lists of tags, whereas t5xxl and Llama (the other two default encoders) are better with natural language.
There is of course the question of weighting across the four encoders, with the current consensus being the Llama encoder is the most heavily weighted. It will take time to understand how best to use the encoders together.
Initially though prompt adherence is good, and what is particularly nice is that making small changes to the prompt does not, in general, reset the whole image so it is easier to evolve an image by tweaking the prompt.
Many different styles can be invoked directly from the base model and I’m sure a list of triggers will appear at some point. Skin tone and anatomy are generally very good but not perfect. The occasional incorrect number of digits, or even entire limbs still make an appearance from time to time.
Sometimes HiDream does seem to get locked onto a particular image composition and changing the seed and tweaking the prompt do not have any great impact - I suspect I need to learn more about the way the four different encoders work.
Overall though I have been surprised how much I have started to use HiDream compared to Flux. The HiDream output has something about it which sets it ahead, not always, but certainly in realistic photographic styles.
HiDream E1
The HiDream E1 model was released even more recently and my experimentation has been more limited. It is touted as a natural language capable image editing model, more aligned to Flux-Dev-Fill for inpainting and outpainting. Again the full model is large so I am using a quantised version. I have used both the ComfyUI published basic workflow and also the HiDream v1.1 workflow by Tenofas which includes an E1 switch.
HiDream E1 is based around image to image with a text prompt without the need for masking to do modifications, however, the input image has to be resized to 768 x 768 and you are advised to keep to those proportions for good results. My limited testing has been mixed and overall disappointing. The results compared to using Flux-Dev-Fill with masked inpainting were poor with some very odd changes being made, along with incorrect blending and lighting. I need to do some more testing at the moment it would not sway me away from the current inpainting and outpainting approach with Flux.
If you cannot run HiDream locally then cloud versions are already available and it is well worth experimenting with.