ComfyUI-CyberMOSS-Sampler

ComfyUI-CyberMOSS-Sampler

<p align="center"> <img src="assets/icon.png" alt="CyberMOSS Sampler icon" width="200"> </p>

A compact MOSS-Image sampler and resolution-aware empty latent node for workflows that already patch the model with LoRAs.

One model connection keeps the model supplied by your LoRA chain active throughout sampling. Choose a preset for your MOSS-Image variant, then adjust any control independently. CyberMOSS uses its own node identifiers and category so it can coexist with CyberKrea Sampler and other sampler packs.

CyberMOSS is the MOSS-Image sibling of ComfyUI-CyberKrea-Sampler. The sampling engine is derived from ComfyUI-KreaPhoton by Kostiantyn Hrytsuk.

Node

CyberMOSS Sampler is under the CyberMOSS category.

Connections

| Option | Description | |---|---| | model | The final MOSS-Image model from your model chain. Connect it after LoRAs. | | positive | Positive conditioning from your prompt. | | latent_image | A 128-channel MOSS-Image / Flux2 latent at 1/16 of the image size, such as the output from CyberMOSS Empty Latent. ComfyUI's own Empty Latent Image also works; it is converted before sampling. | | seed | Controls the initial noise and all seeded sampling-noise streams for reproducible runs. | | negative | Negative conditioning, used when cfg is not 1. Required for the full presets. MOSS-Image's reference uses an empty negative prompt: a text encode node with empty text. Ignored at CFG 1. | | vae | Optional. Fully decodes the first image in the batch and displays a thumbnail on the node. This costs a full VAE decode for that image; the complete LATENT output is unchanged. |

Sampler controls

Selecting a preset immediately fills in all visible control values. You can then change any individual value without losing the rest of that preset.

| Option | Description | |---|---| | preset | Loads turbo, turbo_reference, full or full_reference values, including CFG, layer skipping and the schedule mode. Preview and denoise are preserved. | | steps | Number of denoising evaluations in the sampler loop, from 1 to 100. With CFG above 1 each evaluation runs the model twice. | | sampler | euler is the first-order method used by all presets. euler_2m uses variable-step Adams–Bashforth history on consecutive deterministic steps; set eta0=0 to use it. Restarts and stochastic steps reset its history. | | cfg | Constant classifier-free guidance. 1 for Turbo and DMD models: the negative is not evaluated. 3 is the official value for full models; 3 to 6 can help short prompts. | | skip_layers | Layers skipped in the negative (uncond) pass, such as 22 or 22, 24, 26. Only active when cfg is not 1. Empty disables it. See Layer skipping. | | restart_frac | Fraction of the existing step budget reserved for the restart pass. Active only with schedule_mode=cybermoss and denoise=1. Skipped if the budget cannot fit both descents. | | sigma_r | Noise level to jump back to at the restart. Higher values make the restart stronger. 0 disables restart and keeps the entire step budget for the first descent. | | plunge | Takes a direct end step from sigma 0.75 before the restart refines the image. Meant for few-step models. Active only with schedule_mode=cybermoss and denoise=1. | | detail | Strength of the detail sigma adjustment. Higher values emphasize fine detail; excessive values can look harsh. | | eta0 | Maximum gated ancestral-noise strength during the middle of sampling. 0 disables this extra noise injection. | | sigma_gate | Eta noise is disabled at or below this sigma. Below 0.35, strength ramps up to eta0 at sigma 0.35. A gate of 0.35 or higher uses a hard cutoff. | | contraction | Multiplies the initial noise: 1.0 leaves it unchanged, 0.70 is the turbo value, and 0 removes it. Restart and ancestral noise are separate. Very low values are experimental. | | schedule_mode | cybermoss uses the CyberMOSS schedule with plunge and restart. At full denoise, reference follows the official MOSS training-grid scheduler; simple uses the continuous shift-5 grid for Turbo/DMD workflows. Both ignore plunge and restart. Below full denoise, reference and simple use a continuous schedule tail. Other manual controls still apply. | | denoise | 1.0 generates from noise. Values below 1 refine an existing or upscaled latent using the final requested steps of a longer, plain schedule, without plunge/restart. 0 returns the latent unchanged (an optional VAE thumbnail can still be decoded). | | preview_method | default follows ComfyUI's global preview setting. Explicit overrides are auto, latent2rgb, taesd, and none. TAESD requires taef2_decoder in models/vae_approx. The previous global setting is restored after sampling. |

Presets

MOSS-Image uses a fixed flow shift of 5 for both the full models and the Turbo variant. The presets differ in CFG, step count and sampling extras.

| Preset | For | Steps | CFG | Skip layers | Schedule | Extras | |---|---|---:|---:|---|---|---| | turbo | SFT Turbo, DMD LoRA | 8 | 1 | none | cybermoss | restart 0.25 at 0.65, plunge, detail 0.50, eta 1.0, contraction 0.70 | | turbo_reference | SFT Turbo, DMD LoRA | 8 | 1 | none | simple | neutral | | full | Base, SFT, fine-tunes such as Kromoss | 30 | 3 | 22 | cybermoss | detail 0.30, otherwise neutral | | full_reference | Base, SFT, fine-tunes such as Kromoss | 50 | 3 | 22 | reference | neutral |

full_reference follows the official full-model settings: Euler, 50 steps, CFG 3, shift 5 and layer 22 skipped in the negative pass. At denoise=1, its sigmas match OpenMOSS's MOSSImageScheduler: the shifted inference grid is snapped to the nearest point on the 1,000-step training grid, consecutive duplicates are removed, and zero is appended. At 50 steps the last positive sigma is 0.024.

turbo_reference keeps the continuous simple-style shift-5 schedule at 8 steps and CFG 1. It is a Turbo/DMD baseline, not the full-model reference scheduler; follow your adapter's own sampling instructions when they differ. turbo and full are starting points for tuning, not guarantees. The turbo extras come from CyberKrea's Turbo presets and have not been validated on MOSS-Image checkpoints.

Few-step adapters that ship their own sigmas, such as the 4-step rapid-moss node, should be used with their own sampler and sigma outputs.

Layer skipping

MOSS-Image's reference pipeline skips layer 22 in the negative branch of CFG (uncond_skip_layers=[22], zero-based). The positive pass runs the full model. The user guide also mentions layers 22, 24 and 26 as options that can improve quality.

CyberMOSS applies this through ComfyUI's standard dit / single_block patch-replace hook, only for the uncond pass. LoRAs and other model patches stay active in both passes. It is the same mechanism as ComfyUI's SkipLayerGuidanceDiTSimple node with single_layers set to 22.

Use one or the other. If the model already carries a Skip Layer Guidance cond-batch function, CyberMOSS leaves it in charge, does not apply skip_layers, and logs a warning.

Refinement and latent upscaling

For a second pass, connect an already generated, VAE-encoded or enlarged latent with the same MOSS-Image model and matching positive conditioning. Start with preset=turbo_reference, steps=3 and denoise=0.30 for a Turbo model, or preset=full_reference with around 15 steps for a full model. These are starting points for testing, not a guarantee of image quality. Lower denoise generally preserves more of the source; higher denoise allows more changes. An empty latent has no image to refine.

For 0 < denoise < 1, the schedule uses the final steps transitions from a plain schedule with int(steps / denoise) total steps. This happens in all schedule modes; plunge and restart are bypassed, while detail, eta and contraction remain manually controllable. In reference mode refinement deliberately uses the same continuous tail as simple, since snapping a long expanded schedule to the training grid would merge steps and reduce the requested budget. Denoise is a schedule fraction, not a direct sigma value.

Preset changes preserve denoise. Set it back to 1.0 when returning to full text-to-image generation.

At full denoise, one step always uses the full [1, 0] descent. A restart needs at least two steps without plunge, or three with plunge. Smaller budgets fall back to a single descent while preserving the requested evaluation count.

API

The frontend obtains preset and resolution tables from Python's node metadata; there are no separate JavaScript copies.

ComfyUI API prompts must include the visible required controls. The supplied values take precedence: changing only the preset label does not overwrite them. Headless clients can obtain the templates from GET /object_info/CyberMOSSSampler, at CyberMOSSSampler.input.required.preset[1].cybermoss_presets, and merge their overrides before submitting the prompt. Direct Python callers can use resolve_settings("full", steps=40) to fill omitted values from a preset.

CyberMOSS Empty Latent

CyberMOSS Empty Latent creates the 128-channel latent expected by MOSS-Image (Flux2 VAE, 1/16 of the image size) and outputs the selected width and height as integers. Choose a size tier first; the resolution dropdown then shows only the matching dimensions.

| Option | Description | |---|---| | size | Selects the S, M, L, or XL megapixel tier, filters the resolution dropdown, and preserves the selected aspect ratio. | | resolution | Selects the concrete width, height, and aspect ratio within the chosen tier. | | batch_size | Number of empty latents generated in one batch. Higher values require more VRAM. |

| Aspect | S (~1.0 MP) | M (~1.5 MP) | L (~2.0 MP) | XL (~3.0 MP) | |---|---|---|---|---| | 1:1 | 1024x1024 | 1248x1248 | 1456x1456 | 1776x1776 | | 4:3 | 1152x864 | 1472x1104 | 1664x1248 | 2048x1536 | | 3:4 | 864x1152 | 1104x1472 | 1248x1664 | 1536x2048 | | 3:2 | 1248x832 | 1536x1024 | 1776x1184 | 2160x1440 | | 2:3 | 832x1248 | 1024x1536 | 1184x1776 | 1440x2160 | | 16:9 | 1360x768 | 1680x944 | 1936x1088 | 2304x1296 | | 9:16 | 768x1360 | 944x1680 | 1088x1936 | 1296x2304 | | 21:9 | 1568x672 | 1904x816 | 2240x960 | 2688x1152 | | 9:21 | 672x1568 | 816x1904 | 960x2240 | 1152x2688 | | 5:4 | 1120x896 | 1440x1152 | 1600x1280 | 2000x1600 | | 4:5 | 896x1120 | 1152x1440 | 1280x1600 | 1600x2000 |

Every dimension is a multiple of 16, as MOSS-Image requires. S, M and L follow the official 1, 1.5 and 2 MP recommendation. XL goes beyond it, inside the 4 MP range the user guide describes as extrapolation. The 16:9 and 9:16 buckets approximate their ratio; all others are exact. These buckets have not been validated through generation tests.

Wiring

base model -> LoRA(s) ------------------------> CyberMOSS Sampler (model)
prompt text encode ---------------------------> CyberMOSS Sampler (positive)
empty-text encode (full presets) -------------> CyberMOSS Sampler (negative)
CyberMOSS Empty Latent -----------------------> CyberMOSS Sampler (latent_image)

Installation

Copy the complete ComfyUI-CyberMOSS-Sampler folder to:

ComfyUI/custom_nodes/ComfyUI-CyberMOSS-Sampler

Restart ComfyUI. No extra Python packages are required beyond ComfyUI's own PyTorch/Comfy modules. A ComfyUI version with MOSS-Image support is required. The model check recognizes both ComfyUI's native MOSSImage class and the independent MOSSImage class supplied by the official OpenMOSS custom loaders, including their subclasses.

After updating, also refresh the browser so it loads the matching frontend extensions and node metadata. See CHANGELOG.md for changes.

When upgrading from 0.1.0, reference now selects the official MOSS grid at full denoise. To retain the old continuous reference grid, choose simple. For saved Turbo workflows, select turbo_reference again to load its updated simple setting; loading a workflow preserves its saved manual controls.

Development checks

With Python 3.11+, PyTorch, NumPy and Node.js 22+ installed:

python -B -m unittest discover -s tests -v

Tests cover schedule budgets, tensor-level sampling, seed reproducibility, guidance and layer skipping, previews, validation and frontend widget behavior. ComfyUI boundaries use test doubles; image quality and third-party patch compatibility still need generation tests with a real MOSS-Image checkpoint.

License

Released under the MIT license. The original Kostiantyn Hrytsuk copyright notice is retained alongside Cyberdelia's notice. See LICENSE.