HiDream is a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds.
For more information, please see https://github.com/HiDream-ai/HiDream-I1
Description
hidream_i1_fast_bf16
FAQ
Comments (25)
32GB
Clearly barely an inconvenience for local image generation using an NVidia card that costs more than a replacement kidney for an ailing 8yo.
Onward to the glorious low-cost future where we must eat fried locusts while shivering in cardboard boxes under freeway arches to afford to run our GPU overlords...
On the bright side, freedom of expression (And thought) will likely be outlawed by then, so we won't have much use for generative AI anyways!
@FKAI007 I promise to eat my locusts in silence as a micro-protest.
@FKAI007 freedom of expression will require an encrypted application made by the individual which gets reviewed by AI (the guidelines for the review process are setup by oligarchs) and either approved or denied through an encrypted reply back. "Thought or thinking" on the other hand will not be free-willed, it will be manufactured by the Stargate program and then beamed to neuralink (with guidelines setup up by the oligarchs) which then instructs what your brain will think and it is programmed to make you like it regardless. Not a fantasy but coming soon to a dystopian Orwellian world run by the oligarchs.
@xlr8td The word 'freedom' will be neurologically erased... just as today nobody knows what the well-know word 'pungle' (to delete the wealthy who oppress the poor) means since it was erased from human consciousness in 1973 using the crude satellites of the day and by the careful burning of books from the world's libraries. Nobody even remembers now. It's all very moonwilly and snoorish. I wish it didn't cost me $10 every time I thought a 'bad think'. I may have to sell another of my toes to pay the latest fines. I have non-happy.
id get an elon musk nerolink inplanted in both hemisphers of my brain so i can duel weild AI generation, one half generates a story, the other halfs generates a 4k 60fps HDR dolby atmos (jbl ear drums modification) story. id live 100s of life times in one, my last sight would be a hand with 6 fingers waving bye. a tear runs down my cheek as the nerolink loses driver support and frys my brain
@5090enjoyer For a brief moment, I didn't get the genius comment about 6 fingers, you caught me off guard with that comedy. Don't forget the Psychorama flashes of images sent by the technocrat overlords which will influence you to obey, buy/consume, worship or love/hate.
seems cool but hidream seems to make even 24GB of vram in overheating. plus it doesn't install nodes properly and loads for hours.unusable currently.it needs fixing.I will delete my comment if the issue is fixed but aside from that it looks very promising
I used this set up:
A100 (80 GB) Time: 48.1s for one picture. ;) You won't use this model locally.
https://civitai.com/images/84624143
Would the BF16/FP16 model fit on an RTX 5090?
I tried, it didn't go well. 8 does better.
can this run in forge?
my guess is no, until illysaviel or someone else updates it. it seems to be abandoned sadly. crossing my fingers that im wrong.
just learn comfy. i can teach you if you want!
once you get used to comfy? you'll realize just how capped you were in Forge
It's everything Flux promised to be and more.
it is genuinely so good. The censoring is annoying to the point that it censors even non-explicit stuff, but still. So versatile. It's less good at lengthy prompts, but besides that, the prompt adherence is phenomenal.
I would like to see an equivalent to NoobAI based in this model.
Seeing nobody tried to make a worthy equivalent to Noob based in Flux or SD 3.
People did try with Flux; they just never succeeded.
look into Chroma
Yup, it's Chroma.
this model does not listen: it will fail
a gorilla standing on the top of a building in the bronx throwing apples at the people walking by on the sidewalk below, the gorilla is throwing the apples at a high speed similar to a baseball pitcher with motion blur behind the apples to show the speed, the people down below are throwing the apples back at the gorilla, the people have a apple cannon attached to a world war 2 tank that is painted with wild spray paint colors, the tank is pointed at the gorilla and firing apples at great speed with smoke trails back at the gorilla, the tank is firing at the gorilla, action scene
realistic, actual real world imagery, a crowd made up of, consisting of only cheering Argan Seed humanoids, there are no humans at this exclusive event, Sylvaran Sentinel is at the top of the edge of the ramp grinding the metal edge with the back of the skateboard parallel to the edge of the ramp, the front of the board in a slight upward position also parallel to the edge of the ramp sending sparks flying from the bottom backside of the skateboard that is contacting the metal edge, in a dynamic pose on a skateboard in a skateboarders dream of a combination of skateboard ramps connected in interesting ways, illuminated glowing colorful, concrete, neon glowing spray paint, joyful smiling laughing theme, Argan Seed theme, everything is strategically placed and well thought ou
@mystifying So... HiDream is basically another SD3, Isn't it?
@CuauhtemocI5MAL i havent used that one much, but i have used flux a alot, and flux crea yesterday was the only one to turn a complete crowd into seed people that i randomly tested
@CuauhtemocI5MAL for me, these large models are not more advanced, they are larger databases for the fancy math to extraploate pieces to stick in regions deemed close to the average where they go, when a new wild concept is used then the actual test begins, and they all fail to be a creative ai that is creating.... i cant wait to see a ai master : )