liuliu

27 ms·

421.
by liuliu 4y ago
Yeah, the macOS "native" build was added for "Draw Things", it should support macOS 12.4 and all the way above now.
422.
by liuliu 4y ago
The trick is on how to locate that 0.1% of C++ code ...
423.
by liuliu 4y ago
I think author can try to encode location of the emoji and incorporate better scoring function (GAN)? But that is obviously really "over-engineered" :)
424.
by liuliu 4y ago
TI and hypernetworks (and possibly LoRA) will be doable.
425.
by liuliu 4y ago
I think, we may be able to apply LoRA trick to transfer custom dreambooth models (trained with SD v2) back to depth guide models.
426.
by liuliu 4y ago
Draw Things just added upscaler support (Real-ESRGAN and its ESRGAN derivatives) :)
427.
by liuliu 4y ago
depth2img is underrated for SDv2 release, probably because AUTOMATIC1111 integration is done only 1 or 2 weeks ago. Later this week, Draw Things will have a release with depth2img model as well, initially only supporting iOS-provided depth
428.
by liuliu 4y ago
Since it is a centralized service provided by OpenAI, they should really provide services (paid or not) that given a text, check if it is generated by ChatGPT recently.
429.
by liuliu 4y ago
That's surprising to me, although I did the look about 3 weeks ago, and MPS support is a moving target. It is just M1 without Pro or Ultra right? Also, diffusers does support different backends other than PyTorch.
430.
by liuliu 4y ago
Yeah, it is just PyTorch MPS backend is not fully baked and have some slowness. You should be able to get close to that number with maple-diffusion (probably 10% slower) or my app: https://drawthings.ai/ (probably around 20
431.
by liuliu 4y ago
It doesn't. MPS largely is on GPU. PyTorch's MPS implementation is incomplete a few weeks ago as well. This is about 3x faster.
432.
by liuliu 4y ago
I don't know. This sounds very computer-sciency-ish. We build smaller tools to help build big things. Now the big things are so good and versatile we can replace our smaller tools with the big things too. With the more powerful tools,
433.
by liuliu 4y ago
There are some network changes on the UNet, so if you ported the code over or have mismatched configuration files, it may generate garbage outputs, I wrote some notes here: https://www.reddit.com/r/StableDiffusion/
434.
by liuliu 4y ago
I thought so too until found that there are quite a bit of literatures nowadays about "merging" weights, for example, this one: https://arxiv.org/pdf/1811.10515.pdf and also the OpenCLIP paper.
435.
by liuliu 4y ago
Seems the structure of UNet hasn't changed other than the text encoder input (768 to 1024). The biggest change is on the text encoder, switched from ViT-L14 to ViT-H14 and fine-tuned based on https://arxiv.org/pdf/
436.
by liuliu 4y ago
Bad for tech business whose cost is aggressively on R&D and revenue not inflation-proof.
437.
by liuliu 4y ago
Nobody only spent 10s on linking. For any reasonable sized binary (in the range of 100 to 200 MiB) it is somewhere around 50s to 60s before mold. People regularly links 1GiB binary for living during their development cycle. If you use small
438.
by liuliu 4y ago
"always" is not given. But yes, glad more computing on clients. More power to the people.
439.
by liuliu 4y ago
If you know it will always be short lived, that's fine. But even compilers are not strictly short lived nowadays. Both Swift and Rust (I think) maintains a daemon process to assist incremental compilation.
440.
by liuliu 4y ago
Yeah, I thought Data(contentsOf:) already do that, but it appears not (tested, indeed allocated all the memory to load the data). Adding `mapIfSafe` in the reading options solved this.
441.
by liuliu 4y ago
60s for 30 steps.
442.
by liuliu 4y ago
I use PHPickerViewController: https://developer.apple.com/documentation/photokit/phpickerv... which runs out of the process such that when you want to select a photo into the app, I have no access to any informati
443.
by liuliu 4y ago
I think that fine-tuning the whole model (a.k.a. Dreambooth) on iPhone would require more RAM / processing power than it currently has. More viable path is to implement Hypernetwork + Textual Inversion, that is within possibility of to
444.
by liuliu 4y ago
Should be doable for parameters but at that point, you don't need compression rather just LLM.Int8 tricks would be sufficient. For activations, I wrote about it a while back: https://liuliu.me/eyes/reduce-another-7
445.
by liuliu 4y ago
I have a Twitter thread on this: https://twitter.com/liuliu/status/1587978815208407041?s=46&t... (Note that at that time, there is an implemention bug in inpainting model that caused the weirdness that I need
446.
by liuliu 4y ago
CPU offloading doesn't work because Apple has shared memory arch already. The head slicing is similar to https://machinelearning.apple.com/research/neural-engine-tra... I think it is quite practical only if MPSGra
447.
by liuliu 4y ago
I am using Cloudflare R2, which doesn't have egress fee and I am getting about 5k Class B operations right now. Unless Cloudflare changes their end of the deal, I think it is OK.
448.
by liuliu 4y ago
You should see a warning when selecting that size? 4GiB model cannot run at that resolution until someone implemented FlashAttention on Metal :)
449.
by liuliu 4y ago
Yeah, seems iPad has bunch of issues (curiously most of it related to how I translate tensor back to CGImage ...) stay tuned.
450.
by liuliu 4y ago
Which generation of iPad is that? Honestly I haven't tested iPad yet, still in designing phase.

More ›