Home Creators Posts Import Register Favorites Logout

Videos

  • 2026-10-23 18-54-50.mp4
  • 2026-10-23 19-20-36.mp4

Downloads

Content

You've seen the video: https://www.patreon.com/aifuturetech/posts/fibo-2-0-and-in-173163588 two open-weights models, both happy to edit at around a megapixel. The video stopped there. This is what happens when you push that same edit to 4K.

Two files, one per model. They're separate on purpose. Fibo and MOSS need genuinely different upscale strategies, and welding them into a single graph just means picking the wrong one for one of them.

Same start, different road. Load a reference, write an instruction, get a clean edit at working size. Then Fibo throws the frame away and re-generates the whole thing at 1.5x before handing off to DLSS. MOSS scales with lanczos, runs a light detail pass, then hands over. Both land at 3840 on the long edge, which is 4K on a 16:9 frame.

These add the second pass, the resolution math and the 4K finish.

----------------------------------------

WHAT TO DOWNLOAD (MODEL FILES)

----------------------------------------

FIBO 2.0 WORKFLOW — from https://huggingface.co/Comfy-Org/Fibo-2.0

- fibo2\fibo2_edit_turbo_bf16.safetensors -> models/diffusion_models/fibo2 (I keep the Fibo files in a fibo2 subfolder. Drop them flat if you prefer, then fix the loader path)

- qwen3vl_4b_bf16.safetensors -> models/text_encoders (CLIPLoader type must be "fibo2")

- flux2-vae.safetensors -> models/vae (the full FLUX.2 VAE, not the small one)

No LoRA on this side. Edit Turbo is already distilled, that's the whole point of it.

MOSS IMAGE WORKFLOW — base files from https://huggingface.co/Comfy-Org/MOSS-Image

- moss_image_sft_bf16.safetensors -> models/diffusion_models

- qwen3.5_9b_int8_convrot.safetensors -> models/text_encoders (CLIPLoader type must be "moss_image")

- flux2-vae_small.safetensors -> models/vae (the small VAE here, opposite of the Fibo side)

Direct links, the ones sitting inside the workflow's own Note nodes:

- https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/diffusion_models/moss_image_base_int8_convrot.safetensors

- https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/text_encoders/qwen3.5_9b_w4a8_gptq.safetensors

- https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/vae/flux2-vae_small.safetensors

Now read that twice, because it's the trap in this file. Those note links point at the BASE int8 build and the W4A8 text encoder. The loaders in the graph want moss_image_sft_bf16 and qwen3.5_9b_int8_convrot. Two different pairs living in one workflow. Download what the loaders name, or repoint all three loaders to the int8 pair. Don't pair a base diffusion model with an SFT-tuned text encoder and then wonder why the edits come out vague.

MOSS WORKFLOW — the three LoRAs, all into models/loras/moss/

- moss\Moss-Image-Edit.safetensors — the edit LoRA, strength 1.0. Needs a ComfyUI build with PR #17296 in it, or the edit encoding node won't exist at all. Community source: https://huggingface.co/chinoll/Moss-Image-Edit

- moss\lenovo_moss.safetensors — Lenovo UltraReal, strength 1.0. The realism lifter: https://civitai.com/models/1662740/lenovo-ultrareal

- moss\MOSS_Turbo_LoRA.safetensors — strength 1.0. This is the one buying you a 12-step edit instead of 25 to 30.

Custom nodes: rgthree-comfy for the Get/Set pairs, Labels and the Image Comparer, plus a DLSS 5 pack for DLSS5Settings and DLSS5EnhanceImages, and whatever else your install flags. Fastest way through it: drag the JSON in, let ComfyUI Manager install what it says is missing, restart, reload the file. Don't install anything by hand until you've seen what it asks for.

----------------------------------------

FILES ATTACHED TO THIS POST

----------------------------------------

- Fibo 2.0 Image Edit - 4k Upscale.json (53 nodes — edit, 1.5x re-generate, DLSS 2x)

- Moss Image Edit - 4k Upscale.json (62 nodes — edit, lanczos 1.5x, 0.3-denoise detail pass, DLSS 2x)

----------------------------------------

QUICK START

----------------------------------------

1. Drag in the file that matches the model you want. Start with Fibo if this is your first run. Fewer moving parts, fewer places to get lost.

2. Check three loaders before you queue anything: the diffusion model, the CLIP loader and its type field, and the VAE. That's where almost every broken first run lives.

3. Load your reference. It resizes to 1280 on the long edge by itself, so don't pre-crop to some magic number you found in a comment.

4. Write the instruction in the prompt box. The <image1> convention is covered below.

5. Set a seed. Both files ship with a fixed one so two runs are actually comparable. Switch to randomise once you're happy with the prompt, not before.

6. Queue, then watch the preview. On the Fibo file you get two images: the edit, then the 1.5x re-generate. The second one is the keeper.

7. Your file lands in output/fibo2_edit/ or output/moss_edit/ as a 16-bit PNG in sRGB, at 3840 on the long edge. That's the loop. Nothing else to configure.

----------------------------------------

WHAT TO INPUT (PROMPT FORMAT)

----------------------------------------

Both models want an instruction, not a caption. The one convention that matters is <image1>, which is how you point at the reference you loaded.

Everything else follows from a two-part shape: name the subject the way the model describes it, then close with a preservation clause.

Working example from the Fibo file:

change <image1> female fighter with ponytail, wear yellow track pants with black strap on the side.

Preserving all background, other character, character pose, hair, the same as original.

Working example from the MOSS file:

change <image1> woman wear white bib apron with the Logo "Benji" on the chest area.

preserving all background, character pose, hair, the same as original.

What actually moves the needle:

- Lead with a verb. Change, replace, turn, add. Write a description instead and the model reimagines the whole frame rather than editing one thing in it.

- Keep the preservation clause. It isn't decoration. Drop it and the model starts improving the face, the background and the pose right alongside what you asked for.

- One change per instruction. Ask for two and you'll usually get the second one applied halfway, to the wrong region.

- Name the garment and its colour. "Track pants" gets you a random pair. "Yellow track pants with black strap on the side" gets you those pants.

- Anything that should appear as text goes in quotes, spelled exactly. The attached apron shot is that working: the "Benji" logo came out clean on MOSS, which is that model's home turf.

- Fibo carries a second prompt box on the upscale pass, pre-filled with "Create <image1> high resolution, preserving all background, character pose, hair, the same as original." Leave that wording alone unless you're deliberately chasing a different finish.

----------------------------------------

KEY SETTINGS (AND WHY)

----------------------------------------

FIBO 2.0 WORKFLOW

- Edit pass: 8 steps, CFG 1.0, euler, simple. Edit Turbo is 4-step distilled, so 8 is headroom that lets garment detail land without the face drifting on you. CFG stays at 1.0 because the negative branch does nothing here. Turn it up and you're just buying time.

- No negative prompt node is wired anywhere, and that's deliberate. On a distilled edit model the negative branch is inert. A negative costs you seconds and changes nothing.

- Re-generate pass: second sampler, 8 steps at CFG 1.0, driven by the "Create <image1> high resolution" text. This is a full re-generation of the edit you just approved, not a refine. That's why it holds detail better here than a low-denoise pass would.

- Target resolution: a math node multiplies the resized width and height by 1.5. Your 1280 becomes 1920, so the re-generate pass works at 1920x1080 on a 16:9 frame.

- DLSS 5 stage: 2x (Performance), sharpness 1.5, quality M. 1920 doubled is 3840, and that's your 4K. Leave the other DLSS fields where they are. The preset and sharpness are the only two worth touching.

- Seed 49, fixed as shipped. Change it when you want variety. Changing it because a run looks wrong usually just hides the actual problem.

- Save: 16-bit PNG, sRGB. Keep 16-bit if you plan to grade afterwards.

MOSS IMAGE WORKFLOW

- Edit pass: 12 steps, CFG 1.0, euler, simple. Twelve is where the Turbo LoRA sits comfortably. Below 8 the garment edges soften, above 16 you're paying for steps that don't show.

- Three LoRAs at 1.0 each: Moss-Image-Edit does the instruction following, Turbo buys the step budget, Lenovo UltraReal does skin and fabric. Short on VRAM? Drop Lenovo first. It changes the look least.

- The edit format field on the encoding node has to read "Moss-Image-Edit". That isn't cosmetic. It's what selects the edit instruction template at runtime.

- Negative branch: Conditioning Zero Out, CFG 1.0. Same story as Fibo, the edit path doesn't use a negative.

- Skip Layer Guidance: layer 22, start 0, end 0.5. It's the same SLG that cleans up the free text-to-image workflow. Switch it off for a slightly more photographic read, and pay for it in hair and seam detail.

- First stage comes out around 1280 on the long edge, then a lanczos 1.5x takes it to 1920 before the detail pass. Gentler than sampling straight at 1920, which is the point.

- Detail pass: second sampler, 12 steps, CFG 1.0, denoise 0.3, positive prompt "ultra detail, high resolution, ultra realistic, 8K". Denoise 0.3 is the whole trick. High enough to rebuild pores and fabric weave, low enough that it can't reimagine the edit you just approved.

- DLSS 5 stage: also 2x (Performance), but the tuning slot isn't the same as Fibo's, 0.24 here against 0.15 there. That's per-model tuning, so don't copy one file's values across to the other.

- Seed 42, fixed as shipped.

- One switch node in the Moss file ships off. Leave it off.

----------------------------------------

TIPS

----------------------------------------

1. Do the edit at working resolution, look at it properly, then upscale. The upscale passes are faithful, and faithful means your mistakes get enlarged with everything else.

2. On Fibo the 1.5x pass is a full sampler run, so it will re-interpret small things. If an earring, a logo or a strap survives the edit and then mutates upstairs, don't touch any setting. Repeat the exact noun from your edit prompt inside the "Create high resolution" prompt.

3. On MOSS, upscaling a face-heavy crop, try easing the detail denoise from 0.3 toward 0.4. Past that it starts rebuilding features and the likeness goes.

4. The Image Comparer is already wired A to B in both files. Drag the divider. It's faster than alt-tabbing two windows, and it's the only honest way to tell whether the upscale added detail or just sharpened noise.

5. Keep your source reference at or under 1280 on the long edge before loading it. The resize node does "scale longer dimension" with area sampling, which is good at downscaling. Feeding it a 6K file just costs load time.

6. Only want the edit, not the 4K? Mute the DLSS node and the second save. The edit is already sitting there.

7. These two are deliberately not one workflow. The upscale strategies don't mix. Fibo wants a full re-generate, MOSS wants a low-denoise refine. Merge them and one of the two models is being upscaled the wrong way.

----------------------------------------

COMMON ISSUES

----------------------------------------

Loader goes red on the diffusion model -> you kept the fibo2 or moss subfolder in the path but never made that folder -> create models/diffusion_models/fibo2 (or moss), or strip the prefix from the loader.

Edit does nothing and the output matches the input -> the instruction reads as a caption, or the preservation clause is missing -> start with change/replace/turn and put the clause on its own line.

Fibo comes out washed out and milky -> flux2-vae_small got loaded instead of flux2-vae -> Fibo needs the full VAE, MOSS needs the small one. They do not swap.

Output stops at 1920 -> the DLSS nodes are muted, or the pack isn't installed -> install it, then reload the JSON so the enhance node gets re-instantiated.

The MOSS edit node isn't in the graph at all after loading -> your ComfyUI predates PR #17296 -> update ComfyUI. An older build can't create that node, no workaround.

You clicked the note link and the loader still shows red -> the note links are the base int8 build and the W4A8 encoder, the loaders want SFT bf16 and int8_convrot -> pick one pair, keep it consistent, repoint the other two loaders to match.

Out of memory on the MOSS detail pass -> three LoRAs plus a sampler at roughly 1.9 megapixels -> mute Lenovo UltraReal, or drop the lanczos scale from 1.5 to 1.25 and let DLSS carry more.

The re-generate pass changes the face -> the second prompt describes something the first edit never locked down -> copy the subject noun from your edit prompt into the re-generate prompt word for word.

The 4K PNG is enormous and your editor chokes -> the save node writes 16-bit by default -> go 8-bit for intermediates, and keep 16-bit for the final delivery only.

----------------------------------------

Resources:

- Fibo 2.0 Dev (Bria): https://huggingface.co/briaai/Fibo-2.0-dev

- Fibo 2.0 Edit Turbo: https://huggingface.co/briaai/Fibo-2.0-edit-turbo

- Fibo 2.0 ComfyUI repack: https://huggingface.co/Comfy-Org/Fibo-2.0

- MOSS Image Base (fine-tuning): https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base

- MOSS Image SFT (daily generation): https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT

- MOSS Image ComfyUI repack: https://huggingface.co/Comfy-Org/MOSS-Image

- MOSS Image Edit LoRA: https://huggingface.co/chinoll/Moss-Image-Edit

- ComfyUI support for MOSS Image edit LoRAs: https://github.com/Comfy-Org/ComfyUI/pull/17296

Run the two attached files before you ask me anything about them. Most questions answer themselves somewhere in the first queue.

If something genuinely breaks, drop a comment with your GPU, VRAM, and the exact error and I'll look at it personally.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji

Files

Comments

Pastil

I still don’t get what’s supposed to be good about Moss. It’s decent for non-human imagery, but it’s terrible when it comes to character figures, I honestly can’t figure out what they even managed to pack into a 6GB model file. People just jump on every new thing that comes out like crazy.