FLUX 3 Image Klein Hands-On is CRAZY
Read the Portuguese translation →
Author: Prompt Engineer 48 · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: FLUX 3 Image Klein Hands-On is CRAZY · Channel: Prompt Engineer 48 · 05:13 · Watch on YouTube
Topics: AI content generation, AI infrastructure and hardware, Local and open-source models, Multimodal models
Summary
Section titled “Summary”The video demonstrates running Flux 3 image (called “Flux recline” and “Flex 3 image client” in the transcript) on a rented RunPod A40, covering text-to-image generation, image editing, multi-reference generation, and bounding-box layouts. The setup uses a 5.6-billion-parameter model and a separate prompt upsampler, with the weights requiring substantial disk space. The presenter compares results with the upsampler on and off, and notes that multi-reference results can vary by seed. Generation takes roughly 10 seconds for text-to-image and layout, with higher times for editing and multi-reference; the full GPU session costs about 65 cents.
Key points
Section titled “Key points”- The model has 5.6 billion parameters, and its weights are about 43 GB.
- The presenter runs it on a RunPod A40 with 48 GB of memory, rented for 59 cents per hour.
- The model supports text-to-image, single-image editing, and multi-reference generation with up to 10 reference images.
- The distilled version runs in 16 steps with no guidance.
- The model card advises using a separate 18.8 GB prompt upsampler, which expands a short prompt into a longer scene caption.
- A first run failed with an error involving “apply chat template”; the presenter removed the upsampler block from the pipeline and ran it separately.
- In one comparison, the upsampler produced more image detail and readable text, while turning it off made generation faster.
- Multi-reference results varied: the presenter reports that two of five runs across three prompts came out clean, both with seed one.
- Bounding boxes guide object placement and scale, but are not hard masks.
- Reported generation times were 10.6 seconds for text-to-image, 20.6 seconds for editing, 33.3 seconds for multi-reference, and 9.7 seconds for layout.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:01 | Model overview and requirements | The presenter introduces the model, its supported tasks, the prompt upsampler requirement, and the RunPod A40 setup. |
| 01:04 | Setting up the RunPod environment | The video covers Hugging Face access, downloading the model weights, and troubleshooting the upsampler pipeline. |
| 02:07 | Prompt upsampling and text-to-image tests | The presenter explains the workaround for the upsampler and compares a text-to-image result with the upsampler enabled and disabled. |
| 03:39 | Editing, multi-reference, and layout | The demo edits a bison image, combines bison and shop references, and tests object placement using bounding boxes. |
| 04:41 | Performance, cost, and license | The presenter summarizes generation times, memory use, session cost, and the license condition mentioned in the video. |
Quotes
Section titled “Quotes”5.6 billion parameters, one rented A40 GPU, and about 10 seconds a picture. — 00:01
The distilled version runs in 16 steps with no guidance — 00:33
The upsampler turned it into a 227 word caption in 40 seconds and the image took 9.9 seconds. — 02:07
So, with multi-reference, try more than one seed. — 04:10
The slow part is the swap between the upsampler and the image model, not the image model itself. — 04:41
Concepts and entities
Section titled “Concepts and entities”FLUX 3 Image Klein, Black Forest Labs, RunPod, A40, Diffusers, Gradio, Hugging Face, prompt upsampler, multi-reference generation, bounding box layout
Full transcript (original language)
Section titled “Full transcript (original language)”00:01 5.6 billion parameters, one rented A40 GPU, and about 10 seconds a picture. That is Flux recline, and the whole machine costs 59 cents an hour. All right. In this video, we will install Flux three image client on RunPod and run four things in one small radio app: text to image, and edit a multi-reference shot, and a bounding box layout. Now, mind you, the weights are about 43 GB, which is why I rented a GPU instead of using my laptop. Let’s go ahead and test this out. So, what is it? It is a 5.6 billion
00:33 parameter model from Forest Labs released on October 15th. If you look here on the model page, it does text to image, single image editing, and multi-reference generation with up to 10 reference images. The distilled version runs in 16 steps with no guidance, and the license is free for companies under $5 million in yearly revenue. Now, there is one line on this card that is easy to miss. It says, “Do not use the model without a prompt upsampler.” That is a separate 18.8 GB language model called Klein pup, and it turns a
01:04 short prompt into a full scene caption before the image model sees it. For the machine, I picked an A40 with 48 GB on RunPod secure cloud in the EUS-E1 data center. It is 59 cents an hour. I added a 120 GB disk for the weights, used the RunPod PyTorch 2.8 template, and turned Jupiter on so I can reach a terminal. The model is gated, so you accept the license on Hugging Face and give the pod a token. The token I had saved on RunPod was from last year, and you can see here
01:36 Hugging Face answered invalid username or password. I made a fresh token, and the files came back with access. Then, one script installs the latest diffusers from GitHub and downloads the weights. I skipped the duplicate native checkpoint and the quantized variants, so it pulled 37 GB, and a whole setup was done in about 3 minutes. Now, here’s where I got stuck. The newest diffusers has the upsampler built into the pipeline, so my first run crashed with this error, non type object has no attribute apply chat template.
02:07 What you can do is remove the prompt upsample block from the pipeline and run the upsampler yourself. I keep it on the CPU and move it to the GPU only when it is needed. My first test prompt was seven words, a bison in a misty geothermal landscape. The upsampler turned it into a 227 word caption in 40 seconds and the image took 9.9 seconds. Peak memory was 22.8 gigabytes. So, I wrapped this in a small Gradio app. On the left, there is a prompt box, a place to drop reference images, a
02:37 checkbox for the upsampler, a seed, and a layout box. On the right is the result and below it the exact caption that was sent to the model. First text to image, I type a cozy ramen shop at night in the rain. Keep the upsampler on seed three and click generate. That first wait was about 2 minutes because the model swapped between CPU and GPU, so I sped it up. And you can see the sign reads ramen open late. The menu board is readable and there’s a chef and two customers.
03:07 The image step itself took 10.6 seconds. Now, the same prompt and the same seed with the upsampler off. This one finished in about 10 seconds because nothing has to swap. You can see it is still a good image, a rainy shop front, but the scene is simpler and the sign is not readable English. So, on this one test, the upsampler gave me more detail and readable text. Note the difference between a good image and a broken one. Next, an edit. I drop in the bison image and type same scene at night under a full moon steam lit from behind cool blue light. About 70 seconds later, you
03:39 can see the moon, the blue light, and the same bison standing in the same terrain. The edit step took 20.6 seconds. Now, multi reference. I upload the bison and the ramen shop and ask for the bison standing inside the shop. With seed one, you can see the bison is standing on the wet floor in front of the counter under the warm lamps. This step took 33.3 seconds. But, I want to be honest here. With seed zero, this kind of prompt blended the
04:10 two scenes instead. With geothermal ground showing up next to the teapot and inside the shop, I tried five runs on three prompts and two came out clean, both with seed one. So, with multi-reference, try more than one seed. Last one, a bounding box layout. In the prompt, I tagged the apple and the teapot, and in the layout box, I give each a box on a 0 to 1,000 grid. The apple box is on the left, the teapot box is on the right. You can see the apple lands on the left and the teapot on the right, and it took 9.7 seconds. The
04:41 model card says boxes set placement and scale, not a hard mask, so treat them as guidance. So, here are the numbers from this. A40, text to image, about 10 seconds. Edit, 20.6. Multi-reference, 33.3. Layout, 9.7. Memory stays under 23 GB during generation. The slow part is the swap between the upsampler and the image model, not the image model itself. Cost-wise, the whole session on the A40 from renting the pod to the last recording was about 65 cents of GPU
05:13 time. And remember the license, free under $5 million in yearly revenue, otherwise you need a different license. So, in this video, we have seen how we can run Flex 3 image client on a rented RunPod A40 with text to image, edit, multi-reference, and layout. If you’re interested, you can go ahead and try this on your own pod as well, and let me know how it goes. The model is from Black Forest Labs. I’ll see you in another one. Check out the next video.