Tenos.ai Model Merge Node (Flux) for ComfyUI

A robust, Flux-aware merge node for ComfyUI with deterministic DARE, extra merge modes, a block-weight preset library, regex and model-based masks, safety skip/lock toggles, and an analysis-only dry run. Designed for FLUX-style UNets (and generally compatible with SD-like networks that follow similar block naming).

This node prioritizes predictability, reproducibility, and VRAM sanity. You choose the base model, control per-block weights (or use one-click presets), and get a clean JSON summary with fingerprints so you can re-run the exact merge later.

---

Quick Start

1. Install

   • Copy tenos_merge.py into ComfyUI/custom_nodes/
   • Fully restart ComfyUI (kill the process; don’t just “Reload Nodes”)

2. Add the node

   • Right-click → Add Node → Tenos.ai → Tenosai Merge (Flux+ Deterministic)

3. Wire models

   • model1 = primary
   • model2 = secondary

4. Pick base & mode

   • base_model_choice = the weights you start from
   • merge_mode = algorithm (see “Merge Modes”)

5. Choose weights

   • Pick a preset (block_preset + toggle apply_preset) or
   • Dial block sliders manually (0.0 = keep base, 1.0 = take secondary)

6. (Optional) Safety & Masks

   • Use skip_ / lock_ toggles to protect sensitive parts
   • Use a mask_model or mask_regex to target specific params

7. Run

   • MODEL → merged model (send to KSampler or SaveModel)
   • STRING → JSON summary (audit + reproducibility)

---

Installation & Compatibility

• Put tenos_merge.py in ComfyUI/custom_nodes/, restart ComfyUI

• Node appears as “Tenosai Merge (Flux+ Deterministic)”

• Optimized for Flux-style UNets; generally works with SD-like models using:

  • input_blocks.*, output_blocks.*, middle_block (or mid_block)

---

Inputs (What They Do)

Required

• model1, model2 (MODEL) – the two models to be merged

• base_model_choice (model1 | model2) – which model’s weights you start from

• merge_mode

  • simple – Linear interpolation (lerp)
  • dare – Deterministic DARE: prune smallest diffs, merge the rest
  • weighted_sum – weight_1*model1 + (1-weight_1)*model2 (independent of base)
  • sigmoid_average – Nonlinear remap of amount via sigmoid for softer blending
  • tensor_addition – Explicit delta: t1 + amount * (t2 - t1)
  • difference_maximization – Emphasize where |t2| > |t1|
  • auto_similarity – Cosine similarity per param scales effective blend automatically

Preset Library (Block Weights)

• block_preset (Custom, Balanced, Style-lean, Subject-lean, Text-obedient, Structure-keeper, Detail-boost, Minimal-change)
• apply_preset (bool) — when on, the preset overrides all sliders before merging

What each preset is for

Preset	Intent
Balanced	Neutral 50/50 baseline
Style-lean	Lean into style: middle + upsampling
Subject-lean	Lean into subject/identity: downsampling
Text-obedient	Stronger prompt adherence (conditioning-heavy)
Structure-keeper	Keep composition/structure; conservative everywhere
Detail-boost	Sharpen details; upsampling heavy
Minimal-change	Very light touch; exploratory

Exact preset weights (0.0 keep base ←→ 1.0 take secondary):

{
  "Balanced": {
    "Image Hint": 0.50, "Timestep Embedding": 0.50, "Text Conditioning": 0.50,
    "Early Downsampling (Composition)": 0.50, "Mid Downsampling (Subject & Concept)": 0.50, "Late Downsampling (Refinement)": 0.50,
    "Core Middle Block": 0.50, "Early Upsampling (Initial Style)": 0.50, "Mid Upsampling (Details)": 0.50, "Late Upsampling (Final Textures)": 0.50,
    "Final Output Layer": 0.50, "Other": 0.50
  },
  "Style-lean": {
    "Image Hint": 0.50, "Timestep Embedding": 0.50, "Text Conditioning": 0.50,
    "Early Downsampling (Composition)": 0.30, "Mid Downsampling (Subject & Concept)": 0.40, "Late Downsampling (Refinement)": 0.50,
    "Core Middle Block": 0.80, "Early Upsampling (Initial Style)": 0.75, "Mid Upsampling (Details)": 0.85, "Late Upsampling (Final Textures)": 0.90,
    "Final Output Layer": 0.60, "Other": 0.50
  },
  "Subject-lean": {
    "Image Hint": 0.50, "Timestep Embedding": 0.50, "Text Conditioning": 0.50,
    "Early Downsampling (Composition)": 0.80, "Mid Downsampling (Subject & Concept)": 0.80, "Late Downsampling (Refinement)": 0.70,
    "Core Middle Block": 0.60, "Early Upsampling (Initial Style)": 0.45, "Mid Upsampling (Details)": 0.40, "Late Upsampling (Final Textures)": 0.35,
    "Final Output Layer": 0.50, "Other": 0.50
  },
  "Text-obedient": {
    "Image Hint": 0.25, "Timestep Embedding": 0.60, "Text Conditioning": 0.85,
    "Early Downsampling (Composition)": 0.55, "Mid Downsampling (Subject & Concept)": 0.55, "Late Downsampling (Refinement)": 0.50,
    "Core Middle Block": 0.55, "Early Upsampling (Initial Style)": 0.50, "Mid Upsampling (Details)": 0.50, "Late Upsampling (Final Textures)": 0.50,
    "Final Output Layer": 0.50, "Other": 0.50
  },
  "Structure-keeper": {
    "Image Hint": 0.30, "Timestep Embedding": 0.40, "Text Conditioning": 0.50,
    "Early Downsampling (Composition)": 0.20, "Mid Downsampling (Subject & Concept)": 0.25, "Late Downsampling (Refinement)": 0.30,
    "Core Middle Block": 0.40, "Early Upsampling (Initial Style)": 0.40, "Mid Upsampling (Details)": 0.40, "Late Upsampling (Final Textures)": 0.40,
    "Final Output Layer": 0.40, "Other": 0.20
  },
  "Detail-boost": {
    "Image Hint": 0.50, "Timestep Embedding": 0.50, "Text Conditioning": 0.50,
    "Early Downsampling (Composition)": 0.35, "Mid Downsampling (Subject & Concept)": 0.40, "Late Downsampling (Refinement)": 0.45,
    "Core Middle Block": 0.65, "Early Upsampling (Initial Style)": 0.75, "Mid Upsampling (Details)": 0.85, "Late Upsampling (Final Textures)": 0.90,
    "Final Output Layer": 0.65, "Other": 0.50
  },
  "Minimal-change": {
    "Image Hint": 0.10, "Timestep Embedding": 0.10, "Text Conditioning": 0.10,
    "Early Downsampling (Composition)": 0.10, "Mid Downsampling (Subject & Concept)": 0.10, "Late Downsampling (Refinement)": 0.10,
    "Core Middle Block": 0.10, "Early Upsampling (Initial Style)": 0.10, "Mid Upsampling (Details)": 0.10, "Late Upsampling (Final Textures)": 0.10,
    "Final Output Layer": 0.10, "Other": 0.10
  }
}

Presets don’t bypass safety toggles. If skip_norms is on, those params stay from the base regardless of preset weights.

Block Weights (Manual)

• Image Hint – IP-Adapter / control influence
• Timestep Embedding – how the model interprets noise level
• Text Conditioning – prompt adherence
• Early Downsampling (Composition) – composition & layout
• Mid Downsampling (Subject & Concept) – subject identity & concepts
• Late Downsampling (Refinement) – pre-style refinement
• Core Middle Block – global style / identity
• Early/Mid/Late Upsampling – detail creation & textures
• Final Output Layer – output head / latent projection
• Other – anything not matched by the heuristics

Rule of thumb: Downsampling = what, Upsampling = how it looks, Middle = style core.

Mode-Specific Settings (Plain English)

Context: Every block has a slider (0.0–1.0) used as the blend amount. Some modes add extra knobs. The slider and those knobs work together as described below.

1) simple — linear blend

• What: Straight crossfade between base and secondary.
• Uses slider? Yes (directly).
• Extra knobs: none.
• Use when: Baseline blending; “more of model2 here.”
• Try: 0.2–0.6 on the blocks you care about.
• Pitfall: If models disagree wildly, can look muddy → consider auto_similarity or dare.

2) weighted_sum — global average

• What: final = weight_1*model1 + (1-weight_1)*model2 for every merged block.
• Uses slider? Gate only. Slider > 0 applies the same global mix; 0 keeps base.
• Extra knobs: weight_1 (0..1) — always refers to model1.
• Use when: You want a clean, consistent “70/30” style global blend.
• Try: weight_1 = 0.5 to start; nudge to 0.6–0.8 if base should dominate.
• Pitfall: Sliders don’t shape ratio (beyond on/off). For per-block nuance, use simple/auto_similarity.

3) sigmoid_average — gentler crossfade

• What: Same as simple, but remaps the slider through a sigmoid.
• Uses slider? Yes (curved).
• Extra knobs: sigmoid_strength (higher = steeper middle).
• Use when: Softer, less brittle transitions (style transfer without harsh artifacts).
• Try: sliders 0.3–0.7; sigmoid_strength = 2–4.
• Pitfall: Very high strength “snaps” near the middle.

4) tensor_addition — explicit delta

• What: final = t1 + amount * (t2 - t1) (equivalent to simple, framed as applying delta).
• Uses slider? Yes (linearly).
• Extra knobs: none.
• Use when: You think in terms of “add model2’s change” on top of model1.
• Try: 0.2–0.5.

5) difference_maximization — let the stronger signal win

• What: Where |t2| > |t1|, lean toward model2; else keep model1. Slider controls strength of leaning.
• Uses slider? Yes (strength).
• Extra knobs: none.
• Use when: Push style/details where model2 is clearly stronger without washing out model1 elsewhere.
• Try: 0.2–0.5 on mid/late upsampling; keep skip_norms on.
• Pitfall: Can amplify noise with incompatible models.

6) auto_similarity — smart blend by cosine

• What: Per-param cosine similarity; if tensors are similar → small change; if different → larger change. Effective amount = slider × auto-weight.
• Uses slider? Yes (scales the auto weight).
• Extra knobs: auto_k (higher = more aggressive when difference is high).
• Use when: Great default: preserve stable layers, lean into meaningful differences.
• Try: sliders 0.25–0.6; auto_k = 4–8.
• Pitfall: If models are wildly different, it’ll go big everywhere → lower sliders or auto_k.

7) dare — deterministic DARE

• What:

  1. Compute diff = |t2 - t1|
  2. Prune smallest diffs via quantile (dare_prune_amount)
  3. On the unpruned elements, blend using slider × dare_merge_amount Deterministic: quantile in float32 for stable thresholds.

• Uses slider? Yes (strength on kept elements).

• Extra knobs:

  • dare_prune_amount (0..1): fraction of smallest diffs to ignore
  • dare_merge_amount (0..1): cap/scale of merge strength on kept parts

• Use when: Bring a fine-tune into a base, keeping only the important changes.

• Try: dare_prune_amount = 0.05–0.15, dare_merge_amount = 1.0; normal sliders in the blocks you want.

• Pitfall: Too much prune = “nothing changed”; too little = regular blend. Keep skip_norms on.

Slider interaction cheat sheet

Mode	Does the block slider matter?	How it’s used
simple	Yes	Direct linear blend
weighted_sum	Gate only	> 0 = apply global weight_1; 0 = keep base
sigmoid_average	Yes	Blended through sigmoid curve (softer tails)
tensor_addition	Yes	Add scaled delta (t2 - t1)
difference_maximization	Yes	Strength when (t2 > t1)
auto_similarity	Yes	Scales auto weight from cosine similarity
dare	Yes	Strength only on unpruned elements (after prune)

---

Safety Toggles (Skip / Lock)

Toggle	Protects (keeps base)	Why/When
skip_bias	All .bias	Reduce drift; maintain activation centering
skip_norms	LayerNorm, GroupNorm, BatchNorm (weights + biases)	Preserve calibration; strongly recommended
lock_time_embed	Timestep/time-embedding layers	Keep denoising schedule interpretation stable
lock_conditioner	Text conditioning & related projections	Maintain prompt adherence while changing style/identity elsewhere
lock_output_layer	Final projection / head (e.g., to_rgb, out., final_*)	Avoid breaking the last conversion stage

Tip: Turn on skip_norms by default. Use locks when you want minimum drift in those areas.

---

Masks (Targeted Control)

Semantics: after computing a candidate merged value for a parameter, the mask blends it with the base value:

final = lerp(base_param, merged_param, mask)
# mask = 0.0 → keep base; mask = 1.0 → take merged

Two ways to provide a mask

1. mask_model — a third model whose tensors (by same name) act as masks.
2. mask_regex + mask_value — create a constant mask where the param name matches the regex.

Precedence: mask_model wins for names it covers; regex applies to the rest.

Broadcasting supported:

• exact same shape → used as-is
• [C,1,1] or [C] → broadcast to conv weights [C,H,W,...]
• scalar → broadcast to any shape If a mask can’t be broadcast, it’s ignored (warning in console).

Regex cookbook (Python syntax; escape dots)

• Protect all norms (keep base):

  mask_regex: (?i)(layernorm|groupnorm|batchnorm|ln_|norm)
  mask_value: 0.0
  

• Force final head from secondary:

  mask_regex: (?i)(to_rgb|out\.|final_layer|output_layer)
  mask_value: 1.0
  

• Freeze text conditioning:

  mask_regex: (?i)(text|cond|token|clip|context)
  mask_value: 0.0
  

• Hit only early downsample blocks:

  mask_regex: (?i)^input_blocks\.(?:0|1|2|3)
  mask_value: 1.0
  

---

Execution Controls

• analysis_only — dry run: returns JSON summary; doesn’t modify weights
• calc_dtype — math dtype (float32 or bfloat16). Final params keep original dtype.
• seed — sets torch seeds for deterministic behavior

---

Outputs

• MODEL — the merged model

• STRING — JSON summary including:

  • Sizes (GB) for model1, model2, final
  • Merge mode + full settings (including preset name/applied)
  • Preflight counts (missing in base/secondary, shape mismatches)
  • Run info (seed + key-set fingerprints)
  • Result counts (merged/kept, up to 50 error lines)

Example JSON (trimmed):

{
  "base_model": "model1",
  "sizes_gb": {"model1": 3.21, "model2": 3.21, "final": 3.21},
  "merge_mode": "dare",
  "settings": {
    "preset": {"name": "Style-lean", "applied": true},
    "weights": {"Core Middle Block": 0.8, "...": 0.5},
    "dare": {"prune_amount": 0.1, "merge_amount": 1.0}
  },
  "preflight": {"missing_in_secondary": 0, "missing_in_base": 0, "shape_mismatches": 0},
  "run": {
    "seed": 0,
    "base_keys_fingerprint": "a1b2c3d4e5f6",
    "secondary_keys_fingerprint": "f6e5d4c3b2a1"
  },
  "result": {"merged_params": 1234, "kept_from_base": 56, "errors": 0}
}

---

Block Mapping (Heuristic)

UI Label	Typical param name patterns
Early Downsampling (Composition)	input_blocks.0–3
Mid Downsampling (Subject & Concept)	input_blocks.4–8
Late Downsampling (Refinement)	input_blocks.9+
Core Middle Block	middle_block, mid_block
Early Upsampling (Initial Style)	output_blocks.0–3
Mid Upsampling (Details)	output_blocks.4–8
Late Upsampling (Final Textures)	output_blocks.9+
Final Output Layer	to_rgb, out., final_layer, output_layer
Text Conditioning	text, cond, token, clip, context
Image Hint	ipadapter, control, image.*hint
Other	Anything else

Heuristics are designed to work across Flux and SD-like repos.

---

Recipes

Style over structure

• Base = structure model
• Raise Core Middle Block + Upsampling
• skip_norms ON

Concept blending

• Raise Downsampling (subject/identity lives there)

Auto as smart default

• auto_similarity with moderate sliders preserves stable parts, leans into meaningful diffs

Fine-tune into base (DARE)

• dare_prune_amount 0.05–0.15, dare_merge_amount 1.0
• Adjust block sliders normally

Keep fragile parts steady

• skip_norms and often skip_bias ON
• Lock time_embed, conditioner, output_layer for minimal drift

---

Troubleshooting

• Node not found / old UI appears — You’re loading an older file. Remove duplicates in custom_nodes/, delete __pycache__, restart Comfy.
• Nothing changes — Check block sliders; 0.0 means “keep base”. In DARE, too high prune can mask changes.
• VRAM pressure — Use calc_dtype = bfloat16. Math uses less memory; write-back preserves original dtype.
• Weighted sum feels “different” — By design: weight_1 always maps to model1, independent of base (predictable semantics).
• Import error — Comfy will skip the node. Check the console traceback; fix the file or ping the error.

---

Reproducibility & Licensing

• Set seed for deterministic merges.
• Summary includes key-set fingerprints — if they differ, you’re not merging the same models.
• Respect original model licenses; you’re responsible for redistribution rights.

---

Changelog (highlights)

• Deterministic DARE (quantile in float32)
• Extra modes: difference_maximization, auto_similarity, sigmoid_average
• Skip/lock safety toggles
• Regex + model-based masks with broadcasting
• Preset library with UI controls
• Analysis-only dry run and detailed JSON summary

---