VibeusTwo-stage MOSS-Image-Base at a reported 6 megapixels: Ciclosporin’s... - Vibeus
Two-stage MOSS-Image-Base at a reported 6 megapixels: Ciclosporin’s 16-image gallery pairs a base pass with an 8-step DMD refinement. Portraits, a restaurant, welding and glassblowing scenes, architecture, infrared-style trees, a private jet and a surfer show the range. The creator calls it the current T2I state of the art; this is a personal judgement from selected images, not a controlled comparison. Ten examples are attached here; the original contains all sixteen.
The useful workflow breakdown comes from commenter MastMaithun, rather than the gallery’s creator: first generate with the base model at half the desired resolution, 50 steps, CFG 3, no LoRA, er_sde plus the simple scheduler. Then upscale the latent with nearest to the target size and run a second sampler for 8 steps, CFG 1, with the DMD LoRA at strength 0.9, again er_sde plus simple. Another reply advises lowering denoise for the second pass, without specifying a value. The post does not supply a complete workflow, prompts or seeds, so these settings do not establish how every pictured result was made.
Lodestones’ linked Base adapter card describes a work-in-progress distillation from 50-step, CFG-3 MOSS to eight evaluations without CFG. Its recommended standalone LoRA route differs from that community two-stage recipe: apply lora.safetensors to the stock Base transformer at strength 1, use the usual flow-matching Euler sampler, shift 5, CFG 1 and 4–8 steps. The SFT adapter is a separate download. Neither a full replacement transformer nor the custom data-space sampler is required for the recommended Base LoRA route.
The discussion also points out cherry-picking, skin texture, eyes and an awkward hand in the surfer image. Those examples, including the surfer, remain in this selection. The attached files are Reddit’s 1080-pixel-wide previews converted to JPEG, not the claimed full 6 MP originals.
