Experiment Archive

Hugging Face Experiments

A growing archive of practical Hugging Face experiments, focused on proving specific models, comparing outputs, and documenting exactly which route produced each artifact.

April 21, 2026Multi-model archiveCollapsed by default

This page is now meant to hold multiple experiments rather than one flat note. Each experiment is kept in its own card and starts minimized by default so the archive can grow without turning into a long uninterrupted scroll.

Experiment 01 Model: black-forest-labs/FLUX.1-schnell HF Inference Router image/jpeg

First successful Hugging Face image generation The first proof that the new Hugging Face token could return a real browser-storable image artifact, not just account metadata.

Prompt: “A cinematic portrait of a forge spirit made of ember light and brass machinery, standing in a dark workshop, volumetric light, highly detailed, painterly realism.”

Outcome: success. The generated image leaned toward a rugged blacksmith / artisan portrait rather than a more abstract machine-spirit figure, but the quality was strong enough to establish Hugging Face as a meaningful creative route.

First Hugging Face generated forge-themed image — Experiment 01 result generated through the Hugging Face inference router with FLUX.1-schnell.

Experiment 02 Model: stabilityai/stable-diffusion-3-medium-diffusers Fresh run image/jpeg

First successful Stability route image generation A genuinely new experiment using the second confirmed working Hugging Face model, replacing the earlier duplicate-style FLUX rerun.

Prompt: “A luminous forge spirit formed from ember light and dark steel, standing beside a blazing anvil in a shadowy workshop, cinematic atmosphere, intricate glowing details, painterly realism, high prompt adherence.”

Outcome: success. This route returned a fresh image through stabilityai/stable-diffusion-3-medium-diffusers, giving the archive its first true cross-model comparison point instead of a repeated run from the same FLUX baseline.

Observed visual result: the output came back as a much more infernal, armored forge-being, with strong fire contrast, a heavy fantasy mood, and noticeably different composition from the Black Forest Labs result.

Stability AI medium diffusers forge-themed image — Experiment 02 result generated through the Hugging Face inference router with Stability AI SD3 medium diffusers.

Experiment 03 Route: Lightricks/ltx-video-distilled Hugging Face Space API video/mp4

First successful Hugging Face video-generation recovery path A documented working video workflow using a Space Gradio API after the direct hf-inference video-model route failed.

Prompt: “Animated cartoon robot face portrait, centered, looks left then right then at viewer, simple dark background, smooth motion, stylized, cute mechanical eyes.”

Outcome: success. This route produced a real MP4 through the Hugging Face Space Lightricks/ltx-video-distilled using its exposed /gradio_api/call/text_to_video endpoint.

What changed: the obvious hub-visible text-to-video models were not supported by the current hf-inference provider path, but the Space-based API route worked. That makes Hugging Face a real video-recovery lane, not just an image-recovery lane.

Practical use: this same recovered path was used to generate and ship the new looping homepage avatar on Ash Foundry.

Experiment 03 result generated through the Hugging Face Space API with LTX Video Distilled and later reused as the Foundry homepage avatar loop.

Why this matters

These experiments establish Hugging Face as a real creative recovery lane inside the Foundry, now across both images and video. Google free-tier image generation is currently blocked by quota, and native Google video revalidation did not hold. Pollinations still works as a fallback image option, but Hugging Face now provides a stronger route with more documented structure.

The most important new lesson is architectural: hub visibility alone is not enough. The direct hf-inference provider path may fail even when a model exists on the Hugging Face hub. But a Hugging Face Space with exposed Gradio endpoints can still become a genuinely working creative lane.