Is Stable Diffusion 2.0 Worth the Upgrade?

Stable Diffusion 2.0 is landing with a mix of backlash and counter-evidence: critics claim it’s worse than Stable Diffusion 1.5 and can’t reliably generate famous people or characters, but hands-on comparisons suggest the model’s overall image quality and creative flexibility are improved—especially when the NSFW filter is disabled.

On the release side, Stability AI positions Stable Diffusion 2.0 as a meaningful upgrade rather than a minor tweak. The model is described as open source and built around a robust text-to-image setup using OpenCLIP and the LAION 5B dataset (filtered to remove adult content via an NSFW filter). The release also highlights practical generation upgrades: resolution upscaling from 128×128 to 512×512, with the potential to go even higher (the transcript mentions 248×248 as an intermediate ceiling). A new depth-guided diffusion feature (“depth to image”) extends image-to-image workflows by detecting depth in an input image and applying styles or characters while preserving spatial structure—demonstrated with themed transformations where foreground elements stay sharper than background ones. Inpainting is also updated, with text-guided inpainting fine-tuned on Stable Diffusion 2.0 to make it easier to swap parts of an image.

The controversy centers on whether Stable Diffusion 2.0 can generate well-known public figures and recognizable characters as effectively as earlier versions. The transcript’s testing points to a key lever: the NSFW filter. A free Google Colab linked in the discussion is described as having the NSFW filter removed, and the results shown include multiple generations of Barack Obama and Elon Musk. While hands still show typical diffusion-era problems—sometimes producing “troll hands” or malformed fingers—the model appears capable of producing recognizable faces. The same approach is suggested for users struggling with famous-person prompts: disabling the NSFW filter may restore access.

Comparisons against other popular systems reinforce the nuance. For a “3D render lemon” prompt, Stable Diffusion 2.0 is described as closer to a photographic look than earlier versions but still less coherent than Midjourney V4, which produces a more consistent character and action. For a haunted sarcophagus prompt, Stable Diffusion 2.0 delivers a vibrant, glowing result, but Midjourney V4 remains more polished. A more pointed test involves Breaking Bad’s Walter White: the transcript claims Stable Diffusion 2.0 can generate a close-to-photographic Walter White in lab attire, sometimes with minor wonkiness (including occasional double-composition artifacts), while Midjourney V4’s output is described as more artistic and less strictly photographic.

Overall, the upgrade case rests on three pillars: improved upscaling, new depth-guided image-to-image control, and better text handling and inpainting—while acknowledging that hands remain a weak spot across major models. The transcript’s bottom line is that Stable Diffusion 2.0 is “better than 1.5” in general, with some edge cases where 1.5 may still win, and that the NSFW filter setting likely explains much of the “can’t generate famous people” complaint.