YES SUCCEEDED!!!

Just rendered an image at 944×1152 (slightly above 1024×1024) using Flux1-Schnell-FP8 on my 6700 XT, and it works! (Image 1 is the Real-ESRGAN 2× upscaled version)

Workflow 1: Sampling (Image 2)

Prompt executed → UNet generates the latent

Step 1 (model load + latent generation) took 419 seconds

Output: Latent tensor saved to disk

Workflow 2 : VAE Decode (Image 3)

Latent loaded → VAE decodes the image

Duration: 7.5 seconds

Advantage: UNet doesn’t need to stay in VRAM → VRAM freed, even on 12 GB GPUs

The problem with the stock LoadLatent Node

Dropdown only shows files if they were produced / annotated by a previous SaveLatent Node

Node is designed to pass latents inside a graph, not load arbitrary files from disk

Purpose: prevents accidentally loading wrong files

Workaround (Image 4)

Edited /ComfyUI/nodes.py, class LoadLatent

Hardcoded latent path → Node now loads directly from disk

Result: Workflow 2 runs instantly, UNet can be unloaded

Timing

Step 1 (model load + latent generation): 419 s

Step 2 (VAE decode): 7.5 s

Result: High-res images on a 12 GB RDNA2 GPU are now possible on Flux1-Schnell-FP8 without ComfyUI crashing! (Image 5 is the original output)

This might actually become my new Flux workflow: render quick 512×512 previews first (which works perfectly on RDNA2 GPUs), sort out the good ones, extract the seed from the PNG metadata, and then re-render only the selected images with the same seed using the split workflow at higher resolutions. This way, high-resolution Flux1-Schnell-FP8 renders become possible on 12 GB RDNA2 GPUs D:

Question at the end: Has anyone ever done this before? Because I have no clue xD

#ComfyUI #flux #Flux1SchnellFP8 #FP8 #AMD #RDNA2 #VAE #AIArt #Pixelfed #HighResolution #GPUOptimization #LatentWorkflow #AIWorkflow #AIHacks #RealESRGAN #Upscale #AIExperiment #CreativeAI #DigitalArt #AICommunity #python #linux #opensource #foss

Finally got SDXL running!

I looked into the error
"modules.devices.NansException: A tensor with NaNs was produced in VAE"
and here’s what it means:

Briefly, how an image is generated with a diffusion model: The text encoder interprets the prompt, the UNet "dreams" iteratively in the latent space from noise into an image structure, and the VAE translates this latent vision into visible pixels.

A tensor is simply a multi-dimensional array of numbers, basically the data structure where the model stores all its calculations, like colors, intensities, and intermediate results of the image.

In this case, the VAE experienced a numerical instability: the latent tensor contained invalid values (NaNs), so the dreamed image could not be decoded correctly. In short: the model was still dreaming in the latent space, but the numbers “exploded” along the way (e.g., division by zero, overflow, or undefined operations).

#StableDiffusion #SDXL #AIArt #DiffusionModel #VAE #LatentSpace #Tensor #DigitalArt #CinematicArt #Kunst #KI #AI #DigitalIllustration #StilizedRealism #UrbanFantasy #Motion #DramaticLighting #FilmStill #AICommunity

J'en avais parlé quelques fois: Madame et moi rejoignons la team #velo. 2 #brompton reconditionnés vont bientôt arriver pour faire des balades, remplacer l'auto pour les trajets courts et faire du multimodal les weekends et pendant les vacances.
On est partis sur des #vae parce que nous ne sommes pas des grands sportifs et que la Normandie c'est plus vallonné qu'on ne le pense.
Je vais sûrement avoir plein de questions dans les semaines/mois qui viennent sur les accessoires, les apps GPS...

De passage au Dellec aujourd'hui.
#velo #vae #Finistère #Bretagne

#GoldCurrency - wieviele Länder werden #2026 damit auf die Märkte kommen? Werden #Indien #Brasilien oder #VAE eigene goldgedeckte Währungen einführen?