Image and video generation

Does FLUX.3 Klein run in 8 GB? 5090, 5070 and Spark generation times

The 8 GB support statement applies to the INT8 generation path.

BFL says INT8 FLUX.3 Klein can run on RTX GPUs with 8 GB VRAM¹. The prompt² upsampler is a separate model and uses more memory. Compare published GPU³ generation time separately from your full local workflow.

First pin down settings for 8 GB

FLUX.3 Image Klein is a 5.6B-parameter model that creates images from text or reference images. BFL says its INT8 build runs on GeForce RTX GPUs starting at 8 GB VRAM through ComfyUI or Diffusers. On an 8 GB PC, start with the INT8 distilled checkpoint⁴, 1024×1024, and batch 1. Do not apply this figure to BF16⁵ or other quantized formats. BFL model announcement

ComfyUI’s distilled template defaults to 15 steps and allows 4–16. BFL’s RTX comparison uses 16 steps, so note the step-count difference when comparing it with an unchanged template run. Record both memory and time whenever you change resolution, steps, or batch size⁶.

A monitor beside a PC with a visible dual-fan GPU displays a small image of a blue jug and lemons. — Separate the main model from additional models loaded into GPU memory.

Account for the upsampler’s additional memory

The ComfyUI text-to-image template can use enable_prompt_upsampler to expand a short request into a detailed scene description. If a VRAM error appears while loading the separate Qwen3.5 9B upsampler, turn off that setting and write the subject, setting, composition, and lighting directly in the prompt. ComfyUI template guide

Separate published GPU time from end-to-end time

BFL and NVIDIA’s GPU generation comparison uses distilled, 16 steps, and one image per run. The table separates text-to-image results at 1024×1024 and 2048×2048. ComfyUI defaults to 15 steps, so its unchanged template is not the same setup; API⁷ latency is a separate metric. Published BFL results

BFL and NVIDIA FLUX.3 Klein distilled text-to-image generation times
GPU and implementation	1024×1024	2048×2048
RTX PRO 6000 · INT8 · ComfyUI	3.9 s	12.1 s
RTX 5090 · INT8 · ComfyUI	4.1 s	15.0 s
DGX Spark · NVFP4	7.3 s	43.9 s
RTX 5070 · INT8 · ComfyUI	7.7 s	44.4 s
RTX 4070 · INT8 · ComfyUI	8.9 s	50.4 s

For your own comparison, record the first generation after loading the model separately from later runs. Keep the prompt, resolution, steps, batch, and upsampler setting with the result so different conditions are not mixed. If no image is generated, identify the failing stage and first ComfyUI log error before comparing speed.

Several outputs of the same landscape at different resolutions and sizes — Increasing resolution also increases the number of image pixels to generate.

Choose an order for testing on 8 GB

First save a batch-1 image with the INT8 distilled model. Then compare runs with the prompt upsampler on and off, and increase resolution only if needed. Record whether a PNG was saved, generation time, and the node where a VRAM error occurred. Also note whether you must close other apps for each run to understand available memory headroom.

A work note comparing image generation times beside scene drafts — Record generation time at the same step count and workload.

Terminology notes

VRAM — Memory used by a graphics card’s GPU to store model weights and intermediate values.
Back to the text
Prompt — The input text sent to a model. For generative models it describes the desired result and strongly shapes the output.
Back to the text
GPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the text
Checkpoint — A file containing saved model weights and related state. Versions or tasks in one model family may use different checkpoints.
Back to the text
BF16 — A 16-bit floating-point format for model values. Its exponent has the same width as FP32, giving it a similarly wide value range.
Back to the text
Batch size — The number of inputs or requests processed together in one batch.
Back to the text
API — A defined interface that lets other code call a program’s functions.
Back to the text

Frequently asked questions

What should you check first in this article?: BFL says INT8 FLUX.3 Klein can run on RTX GPUs with 8 GB VRAM. The prompt upsampler is a separate model and uses more memory. Compare published GPU generation time separately from your full local workflow.
How much memory does it need?: FLUX.3 Image Klein is a 5.6B-parameter model that creates images from text or reference images. BFL says its INT8 build runs on GeForce RTX GPUs starting at 8 GB VRAM through ComfyUI or Diffusers. On an 8 GB PC, start with the INT8 distilled checkpoint, 1024×1024, and batch 1. Do not apply this figure to BF16 or other quantized formats. [BFL model announcement](https://bfl.ai/blog/flux-3-image-klein)