Jonathan Fischoff
@jfischoff
Followers
5K
Following
100K
Media
2K
Statuses
26K
Computer person. AI @fal
San Diego, CA
Joined April 2008
This paper is wild. Create a camera with cheap terrible camera lens, but train a diffusion model to recreate a much better image.
50
176
2K
@CartoonsHateHer the power dynamics are messed up because he can collect social security and she can’t.
2
17
956
This is the most exciting paper I've read in a while. Alternate title could have been: "One weird trick to increase your depth map inference 200x.". arxiv: github: Let's go through the details 🧵 1/9
12
130
900
Using Photoshop’s generative fill to extend star wars so it works vertically
31
116
778
@CartoonsHateHer Reminds me when I was dating 12 years ago and I would ask girls who their celebrity crush was. Top answer by a long shot: Jack Black.
16
4
585
“Animate Anyone” was released last night for making pose guide videos. Lets dive in. Paper: Project: .🧵1/
13
117
576
🚨 Major FLUX training update!. A week ago @fal dropped a bombshell with sub-five minute LoRA training. Today we doubled the speed. Same step count, same great quality, but now you can train a LoRA in 2-3 minutes. $2 for ~ 2 minutes. Best deal around.
29
55
444
@kenklippenstein I wish people hated all of it and not just insurance companies. Regulations around hospitals and how they charge aren't great either.
22
1
423
BIG @fal announcement!. Proud to release our new FLUX LoRA trainer Now you can train at higher quality under 5 minutes! Great job @cloneofsimo.
24
53
402
Fantastic result by @kijaidesign using @CiaraRowles1's SVD ControlNet. This is using @Kijaidesign's new SVD CN Comfy node:
3
66
326
🚨 New FLUX trainer release! . Pleased to announce the public release of our FLUX Portrait Trainer. This is a new training method, tuned for portraits, which results in models with improved details, prompt following and better generation of small faces. Link next 👇
12
41
373
@Carnage4Life Same with cars right? The generation that gets the new tech thinks it’s cool. It’s just water for fish if you’re born with it.
6
13
339
@NateSilver538 Everyone keeps taking about the Titanic sinking but why is no one mentioning the hull design issues the Olympic has?.
7
4
325
@HappyFelineMeow @SnarkySteff @fmanjoo @SopanDeb As a society we need to reflect on why we have accepted that having a cop come to your house *increases* your chance of getting shot when it should be the reverse. The safety of cops and civilians is not a zero sum game.
7
53
287
@datepsych Probably more women understand this than men understanding how the older they are the more likely their sperm will have genetic defects.
7
7
276
To show off how fast @fal's real-time LivePortrait advanced mode is, we made this demo. It is a little too fun. Link in next post 👇
13
28
267
34
28
262
Impressed with the quality from "I2V-Adapter: A General Image-to-Video Adapter for Video Diffusion Models" . New way to make train a img2vid model. Looks compatible with AnimateDiff as well. paper: project:
5
52
253
If you’re looking for a way to get started on triton checkout this notebook:
3
33
232
@sdw The answer, is simple: consumers want this. The same reason businesses build anything.
93
2
213
FLUX font style LoRAs are going to be big. Can't believe they work this well
6
37
224
The results in this paper look good. I want to believe.
9
27
219
@moultano Mammals are still tiny (obviously we are influencing the earth just saying the graphic is misleading)
3
13
180
@ArmandDoma the problem with stolen valor is it is a very common attack, so people are kinda numb to it.
6
1
163
15
19
173
New FLUX ControlNets were released by @heyjasperai, and they are supported by @fal's FLUX ControlNet endpoint out of the box. Demo of the upscaler CN. Links to follow. Speed up for the impatient.
4
22
163
Next new FLUX feature on @fal: ControlNets! . Add any FLUX supported ControlNet including ControlNet Union! This works with any LoRA as well. Link in next tweet. Credit for the ControlNets to the InstantX team.
7
13
156
We've shipped a ton of new FLUX features @fal. First up, we've added img2img to our LoRA endpoint. Personalized your img2img now, using an FLUX LoRA.
4
18
152
Tencent released the model and code for their text2video system. There model is very similar to SD but with a 3D U-Net. I was most interested in the design of their U-Net, but there were a ton of details, however, I found this section interesting in light of more recent work.
Latent Video Diffusion Models for High-Fidelity Long Video Generation. github: project page:
5
22
146
I just pushed a new AnimateDiff fine-tune that uses an alternative loss function I'm calling `diff_loss`. Get it here: Why might you want to use a different loss function? . Read the🧵or check out the details in this blog post
11
19
147
@kenklippenstein Even if there was it wouldn’t be that useful. Smart people make terrible decisions all the time because of biases related to their identity.
6
2
149
@HappyFelineMeow @SnarkySteff @fmanjoo @SopanDeb At this point, considering our laws and I doubt what the cop did was against the law, and punishing him would be out of line with other acts we have sanctioned. The fake report should be punished, full stop yes. Cops at your house should not be a death sentence.
4
8
120
@Attractiveincel @CartoonsHateHer that's just what they said my man. you are as surprised as I was 🤷♂️.
0
0
132
Trying out the @fal Gen3 integration. Made a workflow that uses my FLUX dev 80s interior LoRA to make and image for Gen3. Workflow: Results I stitched together with some 80s sounding music.
8
17
138
New video upscaler endpoint on @fal Right now it processes each frame with Real-ESRGAN 2x, but we can extend it to other models. Let me know what we should add as an option next.
15
15
136
Me after reading a twenty page mathematical derivation that results in a one line change to the training code.
10
7
132
Interesting technique to improve text generation. Add a loss to ensure the cross attention maps highlight the text area. Makes me wonder if manually adjusting the attentions maps without training during inference would accomplish a similar thing.
7
26
131
Wrote a blog post on training style LoRAs on @fal. It is really easy to get started. Just upload and go but went over some details and more complex workflows.
8
23
127
New FLUX ControlNet Union was just released. @fal's general purpose Flux+LoRAs+ControlNets endpoint already supported it. Give it a shot here. Major improvement from the alpha CN Union.
3
30
125
@SwannMarcus89 Private unions are fine because the firms will go bk if the unions over play their hand. Not as easy to restart the government.
5
0
125
@RichardHanania George Washington personally wrote a letter to a jewish man in NC to be clear religious freedom met for jews too.
1
0
122
@nntaleb Just be glad you weren’t stabbed. The big plot twist is gonna be when we find out the attacker thought Salman was you.
0
0
117
Reading the SDXL turbo paper: Key idea of Adversarial Diffusion Distillation: you train a network to reconstruct an image but also fool a discriminator. Seems super fast
4
24
118
We at have @fal just launched AuraSR v2! . AuraSR v1 could not upscale JPG compressed images, but v2 does a great job. Blog: Weights: API:
2
23
110
I asked FLUX to imagine luxury homes for different cities using my 80s interior design LoRA. A thread 🧵. Detroit, MI
6
6
112
Boom.
We've raised $23M from Kindred Ventures, Andreessen Horowitz, First Round Capital, Perplexity CEO @AravSrinivas, Vercel’s @rauchg , @balajis and Huggingface CTO @julien_c.
19
1
111
Reading “FreeU” (. The find that scaling the backbone and skip connect features to boost the backbone contribution and decrease the skip-connection low freq contribution has a large improvement on quality. Pretty compelling examples.
5
18
109
Reading through the new Stability video model paper (. I'm just going to comment on the things I find interesting in a 🧵 1/
2
19
107
Doing some LoRA training research with Hunyuan. Might get a usable cake cutting effect out of it.
6
11
108
Horse riding an astronaut has been achieved in discord
8
10
105
@lessin advice I once got about raising: negging doesn’t work on girls but it does with VCs.
1
1
142
Some major FLUX ControlNet improvements from @fal. ControlNets now work for all three generation modes: text to image, image to image and inpainting. Also, XLab's style ControlNets are supported in addition to vanilla CNs and ControlNet Union. Link next post
3
10
101
🚨New Workflow Alert! . Just added CoDef support to Go check it out!
9
18
94
@dieworkwear He would get coffee at my coffeeshop in NYC and I’ve never seen a better dressed man in person.
0
0
86
PSA @fal has two FLUX training options. 1. Our custom fast trainer which takes 2-3 minutes and costs $2. 2. A slower 20 minute trainer based on AI-Toolkit which costs $5. If you want the fast (you do) make sure to choose the first option in the dropdown!
9
9
91
@GarettJones I have a vague recollection of story where Rutherford, who he worked for, pointing out in public that the model was inherently unstable and Bohr had no counter to how the orbits where maintained (because they are not).
2
1
80
I'm not a fan of the "Haskell is faster than C" blog posts. It is almost always optimized Haskell against unoptimized C. Haskell is a very good choice for many problems even if you can't optimize to the level you can with C/C++/Rust etc.
7
9
79
“AVID: Any-Length Video Inpainting with Diffusion Model” extends T2V models like AnimateDiff with inpainting. They also purpose a novel attention guidance mechanism to combine clips. Paper: Project: . 🧵 1/
2
10
86
When you train your LoRA without prior loss preservation.
4
6
85
Arch diagram. Low quality capture is just feed into a ControlNet.
1
1
85
Mark defend yourself with this
7
4
83
Flux Pro requires API access, but here is a ComfyUI node to use it via Fal:
2
9
82
FLUX Multi-LoRA is next level. Here I am combining my personalized LoRA I trained on @fal with @araminta_k's Film Foto LoRA for the style. See following posts for how I trained my personalized .
2
12
82
@sdw I'm not really sure why having the ability to generate a more realistic image is the same as a cancer causing substance. Anyway, for telling stories, making ads etc, you often want realism. I don't think it is that hard to understand.
20
0
80
@dieworkwear Lol. On it. Dress George Hackenschmidt. Measurements: .Height: 5' 9.5".Weight: 204 lb. Neck: 22".Chest: 52".Arms: 19".Forearms: 15.5".Reach: 75".Waist: 34".Thighs: 26.75".Calves: 18". Jesus Christ his neck was insane. Less concerned about that. Everything else.
13
0
80
@samfbiddle @shashj He's using the same definition as in the Geneva convention. It is the difference between dropping on a bomb on a military target (like a bunker with arms) that also has civilians vs a target that only has civilians no military. The targets are different but both are intentional.
6
0
75
@nntaleb Exercise is the stimulus. Recovery generates the benefit. All stimuli are dose dependent.
0
6
77
We've created an optimized endpoint just for loading multiple FLUX LoRAs. Now running at @fal speed here: . Watch how much faster it is compared to our general purpose FLUX endpoint that supports LoRAs and ControlNets. It is also half the price 💸
3
7
76
@dieworkwear Both these guys are under 200 pounds. Show me what to do if you are 200 pounds with the same level leanness.
27
4
73
@10x_er Both suck actually. If there better engineers and not engineering that's bad for the team. If they suck as engineers it's super annoying to work for them.
4
0
72
@RadishHarmers Curious, do you think Israel should attempt to dismantle Hamas? They seem to care more about killing jews more than peace, not exactly sure how that can be tolerated. Things were improving before this, so why attack now? They don't seem like rational actors to me.
56
2
72
Quick little blog post on a trick to adjust the motion of SVD by changing the temporal attention "scale" from `sqrt(n/d)` where `n = 1`, to `> 1` for more motion or `< 1` for less:
2
10
75