Skip to content

AI media harness for planning, generating, and editing locally

TypeScript
20.2 KB
Raw Blame

Generator weights

Settings opens as a full-screen page with sections in the left sidebar. Generator templates open in a popup over the page and save edits immediately. Escape or the popup's close button returns to the generator list. From Settings, Back or Escape returns to the previous screen, preserving project edits.

Video, Image and Music remember separate active generators across navigation and restarts. Browsing saved results or changing scene types does not switch them. The Used seed and Used generator rows describe the selected video take or image result; use the button beside Used generator to select it for future runs.

Reference icon generator chooses the template used for automatic, manually regenerated, and built-in reference icons, independently of the video generator. Only available text-prompt templates appear; Animate cannot generate still icons. If only one is available, it is used automatically. Without a valid saved choice, the video default is used if eligible, otherwise the first available template. Icons use the selected template's steps, attention, MotionCache and LoRAs, including their step overrides.

Sampler and scheduler

Open a generator template's Sampling section to choose its sampler and scheduler. Both start at Model default, which preserves the current checkpoint and adapter recipe. Existing templates keep their behavior.

Generator Samplers besides Model default Schedulers besides Model default
MiniMax H3 Euler, AB2 (CUDA), Re-noising, ER-SDE Beta
MOSS, Fibo 2, Krea 2 Euler, ER-SDE Beta
FastH3 Omni Euler None
Animate Euler Beta when no adapter fixes the schedule
YuE2, MiniMax Music3 None None

Active DMAD and Pika recipes control both choices. These restrictions apply to template, global and scene adapters; saved preferences return when the restriction is removed. Selecting a sampler other than Euler turns off MotionCache. H3 image editing requires Euler or Re-noising. Custom checkpoint metadata is also validated by SlopFab before generation; Beta cannot replace a fixed grid supplied by metadata. An adapter's step-count override alone does not prevent using Beta.

Selections persist in template settings, .slop exports and worker requests. The optional payload fields are sampler (default, euler, ab2, renoise, er_sde) and scheduler (default, beta). Omitted fields mean Model default. Explicit choices require a compatible SlopFab runtime on the generation machine.

Sharing generator templates

Use Export in Settings → Generator to select one or more templates and save a .slop file. Each required model, configured optional file, and referenced LoRA must have an HTTP or HTTPS download URL. The export picker explains which templates need URLs before they can be shared. Downloaded templates retain their source URLs and GPU variants when exported; local paths and download state are excluded.

Use Import to open a .slop file. Templates are added with new IDs; duplicate names receive numbered suffixes. Existing templates, the default generator, and unrelated settings are preserved. Custom LoRAs are included by URL, including their strength, activation, step override and sampling recipe. Matching LoRAs already in the recipient's library are reused. Import does not download weights; use the template's normal Download action when ready.

.slop version 1

A .slop file is UTF-8 JSON, optionally with a BOM. It contains a versioned envelope and typed, independently versioned entries. This separates the file format from individual payloads so future entries can represent project templates, settings bundles, or other shareable data:

{
  "format": "slop",
  "version": 1,
  "items": [
    {
      "type": "generator-template",
      "version": 1,
      "data": {
        "name": "Example generator",
        "modelType": "minimax-h3",
        "defaultSteps": 20,
        "attention": "sage2",
        "mode": "prompt",
        "motionCache": false,
        "sources": {
          "transformer": [{ "url": "https://example.com/transformer.safetensors", "gpuModel": "", "minVramGb": 0 }],
          "textEncoder": [{ "url": "https://example.com/text-encoder.safetensors", "gpuModel": "", "minVramGb": 0 }],
          "videoVae": [{ "url": "https://example.com/video-vae.safetensors", "gpuModel": "", "minVramGb": 0 }],
          "audioVae": [{ "url": "https://example.com/audio-vae.safetensors", "gpuModel": "", "minVramGb": 0 }]
        },
        "loras": [],
        "additionalSafetensors": []
      }
    }
  ]
}

Version 1 supports 1–100 entries, with a 4 MB file limit. The current importer accepts only generator-template entries at payload version 1. Unsupported envelope versions, entry types, payload versions, and malformed entries reject the entire import before settings are changed. Unknown payload fields are rejected so settings from a newer version cannot silently disappear. Future payload changes must declare a supported version and migration explicitly.

Generator payloads carry no machine IDs or paths. sources maps the existing weight roles (transformer, textEncoder, videoVae, audioVae, optional tokenizer) to ordered variant arrays. Each variant has url, optional gpuModel (default empty), and optional minVramGb (default zero). Prompt mode requires the four model roles; Animate requires transformer and both VAEs, plus an additional safetensor with role: "promptEmbedding".

LoRA entries embed name, url, enabled, and strength, plus optional stepOverride and samplingPreset ("dmad-4step"). Their array order is preserved. Additional safetensors carry name, url, and an optional role: "promptEmbedding". Local download and preparation state is never shared. URLs must be HTTP(S) and cannot contain username/password credentials.

LoRAs

Settings → Generators includes a separate LoRAs library. TaoMate 3-Step downloads TaoMate-H3-3step-ComfyUI.safetensors. Turbo downloads minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors and defaults to an editable 4-step override. LightX2V Turbo downloads the silveroxides experimental adapter and defaults to an editable 6-step override and 0.9 multiplier. Hyperflow 8-Step downloads the Hyperflow adapter and defaults to an editable 8-step override for MiniMax H3. These adapters use the same managed weights location, background transfer, progress, cancellation, retry, and Work Queue controls as model weights. Downloads are serialized with weight downloads, and video generation can continue while downloading. Missing managed files become downloadable again. Removing a downloaded LoRA deletes its managed file; Add Lora opens an editor to link a named local file, and removing that library entry leaves the original file untouched.

Open a generator template and use its LoRAs section to add local or downloadable adapters. Downloading the generator also downloads any missing active LoRAs with download links. Already downloaded adapters are reused; disabled and strength-zero adapters are skipped. The generator stays under Download until its model weights and active downloadable LoRAs are available. Progress, cancellation, and retry cover the entire download. If only a LoRA remains, the generator row names it (for example, Needs Turbo LoRA). The editor checks and marks individual model files as Found even while an adapter is missing. Enable any number, adjust their strengths, and move them up or down. Disabled adapters and strength-zero entries are omitted; the remaining paths and strengths are passed to SlopFab in displayed order on both CUDA and Vulkan. SlopFab combines adapter updates by summing them. Negative strengths are supported. An active missing adapter causes an error rather than silently changing the generation.

Open any LoRA from the library to set its Multiplier (default 1). This multiplies its strength in every generator template. For example, a LoRA multiplier of 0.75 and a template strength of 0.5 send 0.375 to the engine. Negative values are supported. A multiplier of 0 disables the adapter, skips its download, and excludes its step override. Existing LoRAs without a saved multiplier keep their original strength.

Open any LoRA from the library to edit its optional Override step count. Any supported scene step count (a whole number from 2 to 2147483647) is accepted. The highest override among enabled, nonzero-strength LoRAs replaces the scene's step count, even when the scene requests more steps. Without an active override, the scene's steps are used. TaoMate defaults to an editable override of 3; older TaoMate schedule settings migrate to this value. All adapters now use the ordinary schedule, whose step count includes the terminal sigma (N steps means N−1 model evaluations). The library and per-template selections persist on this computer; adapter paths are injected into runtime requests and are not saved into portable project files.

Model weights

Generator templates are grouped by model in Settings. The Model selector in each template chooses its model family independently of its name.

MOSS Image generates new images from text on CUDA with Flash attention and defaults to 50 steps. Its download includes the MOSS SFT INT8 transformer, Qwen3.5 9B INT8 text encoder, FLUX.2 small image VAE, and Qwen3.5 tokenizer JSON. It needs no audio VAE. Downloaded MOSS templates appear in the image generator picker and are excluded from video generation. Image editing, Zoom, reference images, character sheets, LoRAs, and MotionCache are unavailable for this model.

The internal capability switches live in src/lib/generatorModels.json, shared by the UI and native request validation. Each model declares imageGeneration, videoGeneration, imageEditing, imageReferences, loras, and motionCache. Change these switches when a model gains support, or add a model entry to define its available operations. Legacy templates without a model family remain MiniMax H3.

Generator templates also offer Enable MotionCache, off by default. It uses SlopFab's default motion-aware denoising reuse settings on CUDA and Vulkan. Reuse can reduce transformer calls but may change detail and motion; short schedules may finish without reuse. It is unavailable in Animate mode. The preference is saved per template and applies to subsequent generation requests. Older runtimes remain usable with it off and report a clear error if enabled without MotionCache support.

The native defaults are threshold 0.15, motion strength 1, warmup 4 calls, at most 2 consecutive skips, active range 0.15–0.95, and subsampling stride 8.

In Settings → Generator, edit a template and enter local paths or HTTP/HTTPS download URLs. Hugging Face blob links are converted to direct file downloads. The bundled First/Last Frame (formerly Minimax H3 Original), References, FastH3 Omni, and Singularity templates each download four model files. References has the same settings, with the minimax_h3_ref2va_pruned_int8_convrot.safetensors transformer in place of the First/Last Frame minimax_h3_fl2va_pruned_int8_convrot.safetensors. FastH3 Omni uses fasth3_omniref_8step_pruned_int8_convrot.safetensors and minimax_h3_video_vae_int8_convrot.safetensors, and defaults to 8 steps. Existing Fast templates receive the updated URLs and step default once, preserving custom settings. First/Last Frame and References default to 20 steps. Singularity defaults to 4 steps and uses Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors on all GPUs, with the same text encoder and VAEs as First/Last Frame. It also includes Turbo, enabled at strength 1 with its 4-step override. Existing Singularity templates receive Turbo once, preserving any existing Turbo selection. All four text encoder downloads use the original Qwen3-VL 32B encoder on all GPUs. The incompatible 4B INT4 ConvRot option is removed from saved Minimax templates; templates using its downloaded file switch back to the original encoder, which may need downloading. The old file is left on disk. First/Last Frame uses Winnougan's minimax_h3_fl2va_pruned-w4a8_convrot_pruned.safetensors fallback at 20 GB or less; References uses the hybrid REF2VA W4A8 model. Both use their respective INT8 transformer above 20 GB; unknown hardware uses the W4A8 fallback. FastH3 Omni uses its Omni transformer on all GPUs, without a W4A8 fallback. Saved Fast templates using the old fallback switch to Omni; an existing Omni download is reused when available. Other local weights and customized download variants are preserved.

Bundled MiniMax weights declare these minimum GPU memory capacities:

Weight Minimum VRAM (GB)
First/Last Frame W4A8 / INT8 12 / 21
References hybrid W4A8 / INT8 12 / 21
FastH3 Omni INT8 22
Singularity INT8 21
Animate INT8 21
H3 Qwen3-VL 32B text encoder 16
H3 video VAE FP16 / INT8 6 / 3
Sol H3 VAE 6
H3 audio VAE FP32 1
MiniMax Music3 transformer INT8 / FP16 / FP32 3 / 5 / 10
MiniMax Music3 text encoder / audio VAE 9 / 1

MiniMax Music3 selects FP32 at 10 GB or more, FP16 at 5 GB or more, and INT8 below 5 GB. Its text encoder still requires 9 GB; smaller GPUs receive the download warning even if the transformer fits. Existing Music3 templates gain the new variants once, preserving downloaded files and customized sources.

Before a generator or shared-component download starts transferring files, Slopus warns if all GPU-compatible variants for any needed weight exceed the detected GPU capacity. The warning lists the GPUs and the unmet minimums, with Cancel and Download anyway actions. Continuing selects the smallest GPU-compatible variant when none fits. A known downloaded dependency is also checked when downloading the rest of a generator. For example, H3's shared 16 GB text encoder still triggers a warning on a 12 GB GPU even though its W4A8 transformer fits. Capacities are not added across GPUs or across model files. Unknown GPU capacity does not trigger a low-memory warning and uses the smallest compatible variant. This checks GPU capacity, not the memory currently free while other apps run. Existing unrestricted bundled sources receive these limits once; custom positive limits, GPU rules, URLs and downloaded paths are preserved.

Templates awaiting downloads appear under Download. Once their weights are downloaded, they move into Generators. Removing or losing their downloaded files moves them back to Download.

Krea 2 Turbo and Krea 2 Raw are text-to-image generators without editing, image references, audio, LoRAs or MotionCache. Each downloads its matching INT8 ConvRot transformer (14 GB minimum VRAM), the Qwen3-VL 4B INT8 ConvRot text encoder (5 GB), and Qwen-Image VAE (0 GB, unrestricted). The tokenizer is embedded in SlopFab. Both use Flash attention on CUDA or portable Vulkan.

The Krea sampling recipe is saved with the template and exported in .slop bundles. Turbo defaults to 8 steps, CFG 0 and sigma shift exp(1.15). Raw defaults to 52 steps and CFG 3.5; its automatic sigma shift depends on the canvas size. Authored CFG, step counts and image sigma shifts override the defaults. Negative prompts affect Krea generation when CFG is greater than zero. Existing catalogs receive the two templates once, preserving customized templates and later removal.

Enable Show advanced options at the end of the template editor, then open Download variants under a weight to add alternative URLs. Each variant can specify part of a GPU model name and a minimum VRAM capacity in GB. Selection prefers a matching GPU-specific variant, then the highest memory tier that fits; ties preserve the listed order. Leave the GPU field blank and VRAM at zero for a universal fallback. Windows hardware detection includes NVIDIA, AMD, and Intel. Memory tiers account for the small amount of VRAM reserved by the driver. If every compatible variant exceeds the detected capacity, downloading requires accepting the memory warning. A zero minimum leaves that variant unrestricted.

Click the template's download icon or Download weights in its editor to download missing weights. Files go into weights beside Slopus.exe. If that location cannot be written, Slopus uses a weights subfolder beside its log file. To use custom locations, edit %LOCALAPPDATA%\Slopus\settings.json, which is created automatically on startup:

{
  "weightFolders": ["D:\\Projects\\weights"]
}

Each entry is an absolute path to the weight folder itself. Slopus checks custom folders in list order, then weights beside the executable, then the log folder's weights directory for an existing completed download. New downloads go to the first writable folder in that same order. These locations apply to both model weights and LoRAs. Changes take effect on the next download or removal without restarting. Startup and the regular availability checks also discover completed model and LoRA downloads in these folders, even when no local paths were saved. An empty list keeps the default locations.

To move previously downloaded weights into a custom folder, move their matching .complete.json records with them so Slopus can recognize and reuse the files automatically. Only files with valid download records can be removed through the app. Other app preferences continue to use WebView2 localStorage.

Downloads continue in the background when Settings is closed. Reopening Settings shows the current progress as a color fill across the generator's template row. The fill tracks completed files plus progress through the current file. Downloads also appear in Work Queue with progress and cancellation controls. They run separately from the generation queue and do not occupy a generation slot. Progress details and cancellation remain available in Settings if the editor is closed. Failed or cancelled transfers keep their downloaded bytes. Choose Retry, or reopen the app and choose Download for the same generator or LoRA, to continue from the saved position. Completed files are reused and progress includes the bytes already downloaded. Downloads do not restart automatically when the app opens.

Partial downloads use a stable .part file and matching .part.json record in the weight folder. Keep both files together if moving an unfinished download. Slopus searches all configured weight folders for existing progress before starting a new file. Partial files cannot be used for generation until complete.

Resume requests use the saved strong ETag with HTTP Range and If-Range, and validate the returned byte range before appending. If the file changed, the server ignores ranges, or no strong ETag is available, Slopus starts that file again instead of combining incompatible data. This follows the HTTP range request rules. No full model file is buffered in memory.

After completion, the path fields show local files, while the source URLs remain stored. Templates needing downloads cannot be selected as default and are omitted from the generation picker. Missing downloaded files are checked at startup, when Settings or the generator page opens or regains focus, and periodically while those pages are open. A missing file restores its URL for redownload.

For a template with download sources, the remove icon deletes its managed weight files and retains the template with its URLs. Manually selected files outside the managed download records are not deleted. Templates sharing a removed downloaded file also become downloadable again. Removing a template containing only manually configured local paths retains the previous behavior of removing the template.