Midjourney web experience is now open to everyone

Post by bambax » Thu, Aug 22, 2024, 3:20 PM UTC

vunderba wrote: post #94

Ever since DALL-E 3 completely eclipsed Midjourney in terms of Prompt ADHERENCE (albeit not quality), I've had very little reason to make use of it. However, in my testing of Flux Dev, I can gen images in roughly 15 seconds in Forge, throw those at a SDXL model such as Dreamshaper in the form of a controlnet and get the best of both worlds, high detail and adherent.

Post by liuliu » Thu, Aug 22, 2024, 4:00 PM UTC

Der_Einzige wrote: post #130

liuliu wrote: post #92

Earlier quoted context omitted.

Quality issue should be mainly due to using FP16 accmulators for GEMM in M1-M2, A14-A16 devices (it is not a problem for SD v1 / SDXL models due to smaller channel count). This is changed to FP32 accmulators for GEMM in these devices with 1.20240820.1 release. q8p should have comparable quality to non-quantized models (in Draw Things, it is called FLUX.1 [dev] (Exact)).

Claims that quantization doesn’t hurt models are made all the time but rely on the fact that almost all evaluations today of LLMs hardly scratch their surface. If we evaluated LLMs properly, even large quants would be detectably worse, and by a significant amount.

Post by Der_Einzige » Thu, Aug 22, 2024, 4:10 PM UTC

liuliu wrote: post #163

Der_Einzige wrote: post #130

Earlier quoted context omitted.

Claims that quantization doesn’t hurt models are made all the time but rely on the fact that almost all evaluations today of LLMs hardly scratch their surface. If we evaluated LLMs properly, even large quants would be detectably worse, and by a significant amount.

A model trained on BF16 that within the range of FP16 have effective bit rate of 13-bit at max (e5m7). Reasonable quantization (at 8-bit) gets you weight error (i.e. L2 distance on weights) down to I think there is a line somewhere between 4-bit to 8-bit that will hurt performance (for both diffusion models and LLM). But I doubt the line is between 8-bit to 13-bit. (Another case in point: you can use generic lossless…

Post by minimaxir » Thu, Aug 22, 2024, 5:58 PM UTC

vunderba wrote: post #111

minimaxir wrote: post #40

Earlier quoted context omitted.

Replicate, although not fully free, has the bonus of outputs from Flux (from their endpoints only) able to be used for commercial purposes whereas that is normally only applicable for Flux schnell: https://replicate.com/black-forest-labs/flux-dev https://replicate.com/black-forest-labs/flux-schnell

Are you sure that's true? https://github.com/black-forest-labs/flux/blob/main/model_li... From the license: "We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs you generate and their subsequent uses in accordance with this License. You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein. You may not use the Output to tra…

Post by Suppafly » Thu, Aug 22, 2024, 7:57 PM UTC

whywhywhywhy wrote: post #157

vunderba wrote: post #94

Ever since DALL-E 3 completely eclipsed Midjourney in terms of Prompt ADHERENCE (albeit not quality), I've had very little reason to make use of it. However, in my testing of Flux Dev, I can gen images in roughly 15 seconds in Forge, throw those at a SDXL model such as Dreamshaper in the form of a controlnet and get the best of both worlds, high detail and adherent.

Dall-E 3 (intentionally) leans away from realism though but in doing so what it leans into is a very tacky and aesthetically naive although competently executed type of image. Gives every image the feeling that you're seeing a bootleg version of a genuine thing and therefore makes everything else it touches feel tacky. Same feeling you get looking at the airbrushed art on a state fairground ride.