Anima Delta Mix
Write into 2.9B / 3.8B insert slots without souping the stem.

IMPORTANT

Donors = 2B only. Never drop 2.9 / 3.8 into donor_stem / donor_gen1 / donor_gen2.

CLIP = Qwen 0.6B only. Qwen 3.5 is not for this node or the baked file.

This is not Remap and not a normal merge. Remap answers which address a tensor should land on when you move 28→40→52. Delta Mix answers what may be written at that address.

The floor map is shared (official 2.9 inserts and the 3.8 trainable_dit_blocks). The write rule is not.

What we refuse

We do not average two checkpoints inside one block. We do not smear a donor across the wrong indices. We do not treat “the last 12 blocks” as the shelf. The stem is the 28 inherited floors. They sit at the start of the tape and in the tail.

What we do

stem     28 inherited slots     freeze (w=0) or lerp
insert   official expansion     insert += r · (donor[src] − kept)
gate     unique skip            drop the whole floor if donor ≈ stem

src is the mapped 2B index, not “whatever sits next to it on a diagram”. An empty niche is better than the wrong floor.


Nodes

DeltaMix-2.9B 40     chassis 40     stem + Gen-1
DeltaMix-3.8B 52     chassis 52     stem + Gen-1 + Gen-2
Lora3.8B_only52      52 only        28/40/52 LoRA onto the stem
                                    Gen-2 left alone by default

Donors are 28-block 2B files. Do not feed 2.9 or 3.8 as a donor. The chassis is already-expanded 2.9 or 3.8.

On 3.8 the two shelves have different jobs: Gen-1 is crop (how it is framed), Gen-2 is surface (a trained file, not the same merge). Gen-2 unique is checked against the stem and against Gen-1. Do not put one branch on both shelves.

How to run it

• Load 2.9 or 3.8 into the matching Delta Mix node.

• Point the file fields at the chassis and the 2B donors on disk (low-RAM path).

• Unique ON. Neighbor / spare / deep OFF unless you are experimenting.

• w=0 leaves the stem as the chassis. w>0 is already a different checkpoint, not “a bit more quality”.

• Bake with ModelSave, not CheckpointSave.

On 3.8, put a 28-LoRA through Lora3.8B_only52 with skip_gen2=true. Stock Load LoRA will drop 28-block indices onto the wrong floors.

A starting point, not dogma:

2.9    inject_delta                 unique 0.05
3.8    Gen-1  inject_delta ~0.35
       Gen-2  reset_muted_then_delta ~0.45
       unique 0.05

How to read the log (short)

stem_slots=…            the 28 floors that are not niches
gen1=…   gen2=…         official inserts
unique src=[…]          donor floors that passed the gate
dropped=['14:vs_stem']  duplicate — that slot is not written
spare wrote=0           spare map off, or it had nothing to do
holes=6                 niches left empty — that is expected

dropped is not a failure. It is the method. If Gen-2 logs vs_D1 on all 12, the second shelf is the same crop as Gen-1: change the donor, do not crank r.

Neighbor spare / deep only show up if you turned the switch on. They are off from the factory.

The point
2.9 already owns ~12 insert floors. 3.8 owns another 12. The node does not mint new billions of weights. It occupies houses that already exist, using a 2B donor, and throws away a floor that adds nothing. The stem is not turned into the average of two schools — so a 28-LoRA after the mix still lands where it would on base.

Why this is not a merge

full-DiT merge / lerp     every block is soup of two files
extend_to_new_layers      LoRA smeared onto inserts as well
Delta Mix                 soup banned in the niche
                          niche gets a delta to kept
                          a duplicate is dropped whole

GitHub - https://github.com/bestluner-create/Anima-Delta-Mix.git